Keywords: Machine Learning and Data Mining, Classification and Clustering
Abstract: This paper proposes a multiple kernel construction method for kernel discriminant analysis. The constructed kernel is a linear combination of several base kernels with a constraint on their weights. By maximizing the margin maximization criterion (MMC), we present an iterative scheme for weight optimization. The experiments on several UCI real data benchmarks show that, the constructed kernel with optimized weights results in high classfication accuracy, compared with multiple kernel learning under the framework of support vector machines. The experiments also show that the constructed kernel relaxes parameter selection for kernel discriminant analysis to some extent.

Keywords: Gesture and Behavior Analysis, Pattern Recognition for Surveillance and Security
Abstract: Falls and fall-related injuries remain a major problem in public health domain. For delivering instant and adequate medical support, the automatic, reliable and immediate detection of falls is important。 In this paper, we present a new real-time weightlessness-like-status based fall detection (WFD) method using a single tri-axial accelerometer attached at waist or in trousers pocket. The accelerometer signals are analyzed by vertical direction calibration, threshold detection, and weightlessness-like status detection to discriminate falls from normal daily living activities. The result shows that this algorithm can detect falls with robust sensitivity of 96.5% and specificity of 95%-100% respectively for different accelerometer locations.

Keywords: Pattern Recognition for Surveillance and Security, Pattern Recognition for Search, Retrieval and Visualization, Statistical, Syntactic and Structural Pattern Recognition
Abstract: In this paper a new block-based algorithm has been proposed to deal with facial occlusion when only one sample per person is available. A Local Binary Pattern (LBP) descriptor is applied on the image sub-blocks to extract distinctive texture features from those areas separately. Chi-Square is employed as histogram similarity metric in local classifiers corresponding to different image blocks. Finally, a weighted majority voting scheme is used for decision fusion. Local entropy is proposed to devote weights to classifiers results according to the block informative richness. This way, we can reduce the effect of blocks with appearance deformation on the final decision. Experimental results show the significantly high recognition accuracy of our method on the challenging AR face database compared to recent well-known approaches, without imposing computational complexity.

Keywords: Biometrics, Classification and Clustering, Machine Learning and Data Mining
Abstract: In this paper, we extend the idea of sparse representation into the high dimensional feature space induced by the kernel function, and propose a kernel based test sample sparse representation and classification algorithm (KTSRC) for the first time. The KTSRC is based on the assumption that the test sample can be linearly represented by a part of the training samples in the high dimensional feature space. Although the explicit form of the sample in the feature space is unknown, we can implement the KTSRC by the kernel trick. The experimental results show that the KTSRC achieves promising performance in face recognition, and outperforms the state-of-the-art methods.

Keywords: Classification and Clustering, Biometrics
Abstract: Occlusion problem is one of remaining challenges in face recognition. This work expresses an occluded image as the summation of a non-occluded image and a sparse occlusion. By solving a norm minimization problem, we isolate the sparse occlusion from the face image, and simultaneously reconstruct the image. The reconstructed image is same to the original one in most pixels. To classify an occluded image with unknown identity, we first linearly express it using the images of every person, and then make decision based on the residues of the expressions. This paper also presents the relationship between the proposed method and the popular methods. The experiments validate the feasibility of the proposed method.

Keywords: Classification and Clustering, Detection, Separation and Segmentation, Image and Video Processing
Abstract: A robust and efficient skin detector facilitates automatic human detection and tracking. In this paper, we propose a new skin detection method via linear regression tree, which decomposes the problem of discriminating different skin and nonskin colors into several simple problems. Experimental results on the MCG skin database demonstrated its better generalization ability and discriminability than state-of-the-arts.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Non-myopic active learning allows the learner to select multiple unlabeled samples at a time. It avoids tedious retraining with each selected sample, and is effective to utilize multiple labelers. But current non-myopic active learning methods are typically greedy by selecting top N unlabeled samples with maximum score. While efficient, such a greedy active learning approach cannot guarantee the learner's performance. In this paper, we introduce a near-optimal non-myopic active learning algorithm that is efficient and simultaneously has a performance guarantee. Our experimental results on UCI data sets and a real-world application show that the proposed algorithm outperforms the myopic active learning method and the existing non-myopic active learning methods in both efficiency and accuracy.

Keywords: Feature Reduction and Manifold Learning, Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: In this paper we address the robust face recognition problem for color faces with large variations in pose, illumination and facial expression. A novel algorithm is proposed, namely the Multilinear Color Tensor Dis- criminant (MCTD) model. This approach utilizes tensor representation to preserve image structure, as well as enhance discriminate capability via color space trans- formation. On the other hand, it uses the multilinear analysis technique to handle variations in pose, illu- mination and expressions and improve the performance via minimizing the least square of reconstruction error in the tensor framework. Extensive experiments con- ducted on the CMU-PIE and CurtinFaces databases demonstrate the effectiveness of the proposed approach.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Pattern Recognition for Search, Retrieval and Visualization
Abstract: Statistical mixtures such as Rayleigh, Wishart or Gaussian mixture models are commonly used in pattern recognition and signal processing tasks. Since the Kullback-Leibler divergence between any two such mixture models does not admit an analytical expression, the relative entropy can only be approximated numerically using time-consuming Monte-Carlo stochastic sampling. This drawback has motivated the quest for alternative information-theoretic divergences such as the recent Jensen-R'enyi, Cauchy-Schwarz, or total square loss divergences that bypass the numerical approximations by providing exact analytic expressions. In this paper, we state sufficient conditions on the mixture distribution family so that these novel non-KL statistical divergences between any two such mixtures can be expressed in generic closed-form formulas.

Keywords: Pattern Recognition for Surveillance and Security, Detection, Separation and Segmentation, Segmentation, Color and Texture
Abstract: This paper presents a real-time smoke detection algorithm. A modified Center Symmetric Local Ternary Pattern (CS-LTP) is proposed as the smoke texture descriptor. The change in background texture provides a means to differentiate between smoke and non-smoke region. Combined with the color information of smoke, our method is able to achieve real-time performance at minimum of 30 fps. The comparison with the state of art is presented as well. The experiment results show the proposed method can obtain good performance.

Keywords: 2D/3D Object Detection and Recognition, Image and Video Understanding, Scene Understanding
Abstract: In this paper we present a hybrid generative-discriminative approach for image categorization in real-world images, based on Latent Dirichlet Allocation and SVM classifiers. We use SVMs with non-linear kernels on different visual features in a multiple kernel combination framework. A major contribution of our work is also the introduction of a novel dataset, called MICC-Flickr101, based on the popular Caltech101 and collected from Flickr. We demonstrate the effectiveness and efficiency of our method testing it on both datasets, and we evaluate the impact of combining image features and tags for object recognition.

Keywords: Classification and Clustering, Feature Reduction and Manifold Learning, Machine Learning and Data Mining
Abstract: The complex movement sequences of golf require supporting tools for players and coaches alike. We developed a system that classifies the experience level and trained it with data from an inertial sensor on the club head. Based on 315 golf putts from eleven subjects the system differentiated between experienced and unexperienced players with a classification rate of 86.1%. To improve the classification system and obtain discriminant features we additionally integrated a feature selection step. We compared different selection approaches and concluded that a leave-subject-out feature selection was the appropriate approach to predict the true performance of a live system. The selected features can be fed back to coaches and help them to guide players to a better putting technique.

Keywords: Biometrics, 2D/3D Object Detection and Recognition
Abstract: We present a real-time 3D face identification system using a consumer level depth camera (PrimeSensor). Our system takes a noisy sequence as input and produces reliable identification. Instead of registering a probe to all instances in the database, we propose to only register it with several intermediate references, which considerably reduces processing, while preserving the recognition rate. The presented system routinely achieves 100% identification rate when matching a (0.5-4 seconds) video sequence, and 97.9% for single frame recognition. These numbers refer to a real-world dataset of 20 people. The methodology extends directly to very large datasets. The process runs at 20fps on an off the shelf laptop.

Keywords: Biometrics, Pattern Recognition for Surveillance and Security, Pattern Recognition for Bioinformatics
Abstract: This paper presents an effective spectral correspondence method for fingerprint matching. Minutia matching is formulated as recovering the dense sub-block in the corresponding matrix. Then the spectral correspondence method is used for searching the dense sub-block. First, we propose the pairwise adjacency matrix (PAM), whose diagonal elements represent similarities of local minutia structures and other elements represent pairwise compatibilities between local minutia structure pairs, thus similarities information and compatibilities information are unified appropriately. Second, correct minutia pairs are likely to establish both large similarities and large compatibilities among each other and they form a dense sub-block. Then minutia matching is formulated as recovering the dense sub-block in the PAM. It gives a clear mathematical meaning for "optimal matching minutia pairs". Third, we recover the dense sub-block based on spectral correspondence method, by using the principal eigenvector of PAM and imposing the one-to-one mapping constraints. Proposed method has stronger description ability and better robustness. Experiments conducted on FVC database demonstrate the effectiveness and the efficiency.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining, Classification and Clustering
Abstract: Graph-based methods are an important category of semi-supervised learning techniques. However, in many situations the graph representation of relational patterns can lead to substantial loss of information. This is because in real-world problems objects and their features tend to exhibit multiple relationships rather than simple pairwise ones. In this paper, we develop a semi-supervised learning method which is based on a weighted hypergraph representation. There are two main contributions in this paper. The first is that we develop a hypergraph representation based on the attributes of feature vectors, i.e. a feature hypergraph. With this representation, the structural information latent in the data can be more effectively modeled. Secondly, to address semi-supervised classification, we derive a ell_{1}-norm for the spectral embedding minimization problem on the learned hypergraph. This leads to sparse and direct clustering results. We apply the method to the challenging problem of gender determination using features delivered by principal geodesic analysis (PGA). We obtain a classification accuracy as high as 91% on 2.5D facial needle-maps when 50% of the data are labeled.

Keywords: Feature Reduction and Manifold Learning, Machine Learning and Data Mining, Pattern Recognition for Surveillance and Security
Abstract: Canonical Correlation Analysis (CCA) is a powerful technique for finding the correlations between two sets of multidimensional variables. Due to its performance in practice, many extensions were brought forward such as least square CCA. However, there is not a unified solution to compare their performance, i.e. in the sense of extracting canonical correlations. In this paper, we propose a framework to systematically evaluate performance of CCA and its variants. Firstly, a Correlation Generation Model (CGM) to analyze CCA in three aspects: 1) Why are the multidimensional variables correlated? 2) How are they correlated? 3) How to evaluate this correlation? Based on CGM, it is possible to qualitatively study CCA in terms of accuracy and robustness. Most interestingly, the analysis reveals that CCA actually suffers from the Under Sample Problem (USP), which is often discussed in the machine learning field but ignored in the literature. Finally, experiments based on CGM are performed to evaluate the CCA as well as its variants.

Keywords: Feature Reduction and Manifold Learning, Machine Learning and Data Mining, Classification and Clustering
Abstract: Extracting low-dimensional structures from high-dimensional space through spectral analysis has been prevalent in the fields of machine learning and computer vision. However, most manifold learning methods assume that there is a dominant low-dimensional manifold, while other variations are usually considered as noise or even ignored. This paper proposes a novel submanifold decomposition (SMD) algorithm, which simultaneously considers multiple submanifolds intertwined in the same high-dimensional space for decomposition. It makes full use of multi-category labels of a dataset to improve the modeling of manifolds of each label. In order to applied the proposed method to practical applications, the linear version of SMD is developed subsequently. Comparative experiments demonstrate that the proposed method not only effectively extracts submanifolds by subspace learning, but also outperforms traditional manifold and subspace learning methods for visual recognition tasks.

Keywords: Gesture and Behavior Analysis, Motion, Tracking and Video Analysis, Human Computer Interaction
Abstract: A ceiling sensor system reported in this study is to recognize different activities of multiple persons in the home environment. The sensors output binary sequences by which we know the existence/nonexistence of persons under the sensors. A short-period average of the binary response is shown to be regarded as a pixel value of a top view camera, but the camera-like view is more advantage in the sense of preserving privacy. Using the "pixel values" as features, support vector machine (SVM) classifier succeeded to recognize eight activities of five subjects at average recognition rate of 80.10%. This accuracy is not sufficient in general but surprisingly high with such low-level information.

Keywords: Pattern Recognition for Surveillance and Security, 2D/3D Object Detection and Recognition
Abstract: We propose a generic framework to handle missing weak classifiers at prediction time in a boosted cascade. The contribution is a probabilistic formulation of the cascade structure that considers the uncertainty introduced by missing weak classifiers. This new formulation involves two problems: 1) the approximation of posterior probabilities on each level and 2) the computation of thresholds on these probabilities to make a decision. Both problems are studied and solutions are proposed and evaluated. The method is then applied on a popular computer vision application: detecting occluded faces. Experimental results are provided on classic databases to evaluate the proposed solution related to the basic one.

Keywords: Pattern Recognition for Bioinformatics, Pattern Recognition for Search, Retrieval and Visualization, Classification and Clustering
Abstract: Structural biology is a branch of life science concerned with the study of the structure of biological macromolecules like proteins. The structure of a protein gives much more insight in its functions than that of its amino acid sequence. Protein structure comparison is important for understanding the evolutionary relationships among proteins, predicting protein functions, and predicting protein structures from the chemical composition. In this paper we propose a new approach for structural block retrieval based on the Generalized Hough Transform (GHT). A first technique uses as primitives the single Secondary Structure (SS), an alternative adopts co-occurrence of SSs couple, the third approach uses SSs triplets, and finally the primitive can be an entire block. In this paper we describe some experiments for the retrieval of elementary structural blocks consisting of four- and five-SSs.

Keywords: Gesture and Behavior Analysis, 2D/3D Object Detection and Recognition, Feature Reduction and Manifold Learning
Abstract: We describe a mechanism based upon activity manifolds that map image data from more than one view to spatial pose. We learn the manifolds from training data which are motion capture data about real human subjects exercising the target actions. The nature of the training data allows the learned manifolds to conform naturally to multiple constraints, including (1) the body-part articulation constraint; (2) the image-consistency constraint; and (3) conformation to prior information about the possible human activities. A mirror function is specifically designed that allows the system to pick up the proper manifold in multi-activity scenario. Human pose in both the image space and 3D is expressed in terms of the body-part joint positions. Such a representation allows image data to be related across views and to 3D space with ease. Experimental results show that not only do the manifolds effectively map image data to 3D pose; the presence of multiple images also improves the precision of the recovered pose and helps fix feature extraction error in any single image.

Keywords: Pattern Recognition for Bioinformatics, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: Chemoinformatics techniques aim to predict molecule�s properties through informational tech- niques. Computer science�s research fields concerned by chemoinformatics are machine learning and graph theory. From this point of view, graph kernels provide a nice framework combining these two fields. We present in this paper two contributions to this research field: a graph kernel based on an optimal linear combination of kernels on acyclic patterns and a new kernel on the cyclic system of two graphs. These two extensions are validated on two chemoinformatics datasets.

Keywords: Gesture and Behavior Analysis, Image and Video Understanding, Motion, Tracking and Video Analysis
Abstract: Early Recognition of human activities is a highly desirable functionality for many visual intelligent systems. However, in computer vision, very few work have been devoted to this challenging and interesting task. In this paper, we address human activity early recognition as a pattern recognition problem of time series data. A new model called ARMA-HMM is introduced to integrate both the predictive power of sequential model HMM and time series model ARMA. We also present a novel feature called Histogram of Oriented Velocity (HOV) to encode activity video as a sequential observation of motion signals. Experiments on a daily activity dataset and a realistic YouTube sports dataset show promising results of the proposed method.

Keywords: Classification and Clustering, Image and Video Understanding
Abstract: This paper proposes a cascaded classifier framework for better image recognition. The proposed method is based on the confidence values given by the classifiers. By using our proposed topN-Exemplar SVM in the second stage and comparing the confidence values with those from the first stage, the classification results with less confidence are successfully updated. The validity of our algorithm has been demonstrated by the experiments using three standard image datasets.

Keywords: Machine Learning and Data Mining, Feature Reduction and Manifold Learning, Classification and Clustering
Abstract: Most existing feature selection methods focus on ranking individual features based on a utility criterion, which neglecting the correlations among features. To overcome this problem, we develop a novel feature selection technique using the spectral data transformation and by using ell_{1}-norm regularized models for subset selection. Specifically, we propose a new two-step spectral regression technique for unsupervised feature selection. In the first step, we use kernel entropy component analysis (kECA) to transform the data into a lower-dimensional space so as to improve class separation. Second, we use ell_{1}-norm regularization to select the features that best align with the data embedding resulting from kECA. The advantage of kECA is that dimensionality reducing data transformation maximally preserves entropy estimates for the input data whilst also best preserving the cluster structure of the data. Using ell_{1}-norm regularization, we cast feature discriminant analysis into a regression framework which accommodates the correlations among features. As a result, we can evaluate joint feature combinations, rather than being confined to consider them individually. Experimental results demonstrate the effectiveness of our feature selection method on a number of standard face data-sets.

Keywords: Pattern Recognition for Bioinformatics, Statistical, Syntactic and Structural Pattern Recognition
Abstract: The aim of this paper is to introduce a classical method of pattern recognition as the solution for the medical imaging, and to provide a new angle of using the pattern recognition theory for MEG source reconstruction. We explore a new method of MEG source spatio-temporal reconstruction based on modeling the neural source with extended basis functions. Inspired by the graph theory that Laplacian eigenvectors of spherical mesh are equivalent to its basis functions representing the cortex mesh, we build a new model to describe the current source distributed on each mesh vertex. This model consists of analogous basis functions and unknown weighted coefficients. Along with leadfield, the weighted coefficients can be calculated in the light of the forward formulae of MEG. Expanding this process from a single time point to continuous time series, it is able to obtain the spatio-temporal reconstructed source distributed on cortical mesh vertices. Under the condition of zero-mean Gaussian noise with small value of variance, the results show robustness to noise and better performance than minimum-norm, but intensive to the deep sources.

Keywords: Biometrics
Abstract: In order to achieve high accuracy of face recognition, detection of facial parts such as eyes, nose, and mouth is essentially important. In this paper, we propose a method to detect eyes from frontal face images. The proposed method consists of two major steps. The first is two dimensional Hough transformation for detecting circle of unknown radius. The circular Hough transform first generates two dimensional parameter space (x, y) using the gradient of grayscale. The radius of circle r is determined for each local maximum in the (x, y) space. The second step of the proposed method is evaluation of likelihood of eye using histogram of gradient and Support Vector Machine (SVM). The eye detection step of proposed method firstly detects possible eye center by the circular Hough transform. Then it extracts histogram of gradient (HOG) from rectangular window centered at each eye center. Likelihood of eye of the extracted feature vector is evaluated by SVM, and pairs of eyes satisfying predefined conditions are generated and ordered by sum of the likelihood of both eyes. Evaluation experiment is conducted using 1,409 images of the FERET database of frontal face image. The experimental result shows that the proposed method achieves 98.65% detection rate of both eyes.

Keywords: Classification and Clustering, Machine Learning and Data Mining, Human Computer Interaction
Abstract: Prognosis refers to the prediction of the future health status of a patient. Providing prognostic insight to clinicians is critical for physician decision support. In this paper we present a collaborative disease prognosis strategy leveraging the information of the clinically similar patient cohort, using a Local Spline Regression (LSR) based similarity measure. To improve the reliability of the approach, the algorithm can also incorporate physician�s feedback in the form of whether the patients in a retrieved cohort are indeed similar to the query patient. The proposed methodology was tested on a real clinical data set containing records of over two hundred thousand patients over three years. We report the retrieval as well as prognosis performance to demonstrate the effectiveness of the system.

Keywords: Machine Learning and Data Mining, Classification and Clustering
Abstract: In this paper, we present a theoretical analysis on learning anchors for local coordinate coding (LCC), which is a method to model functions for data lying on non-linear manifolds. In our analysis several local coding schemes, i.e., orthogonal coordinate coding (OCC), local Gaussian coding (LGC), local Student coding (LSC), are theoretically compared, in terms of the upper-bound locality error on any high-dimension data; this provides some insight to understand the local coding for classification tasks. We further give some interesting implications of our results, such as tradeoff between locality and approximation ability in learning anchors.

Keywords: Gesture and Behavior Analysis
Abstract: The use of wearable sensors for human activity monitoring and recognition is becoming an important technology due to its potential benefits to our daily lives. In this paper, we present a sparse representation-based human activity modeling and recognition approach using wearable motion sensors. Our approach first learns an overcomplete dictionary to find the motion primitives shared by all activity classes. Activity models are then built on top of these motion primitives by solving a sparse optimization problem. Experiments on a dataset including nine activities and fourteen subjects show the advantages of using sparse representation for activity modeling and demonstrate that our approach achieves a better recognition performance compared to the conventional motion primitive-based approach.

Keywords: Biometrics, Pattern Recognition for Bioinformatics, Handwriting Recognition
Abstract: This conference paper investigates the possibility of using on-line handwritten signatures for biomedical biometry. More specifically, features extracted from sigma-lognormal representations of signatures are applied to the problem of brain stroke susceptibility assessment. The area under the receiver operating characteristic curve (AUC) is used to evaluate the predictability of the most important modifiable brain stroke risk factors (diabetes, hypertension, hypercholesterolemia, obesity, cigarette smoking, cardiac problems) based on four different statistical modeling of the features� variation (random forest, linear discriminant analysis, logistic regression and linear regression). Our preliminary results show a potential predictability (AUC of about 0.7-0.8) for every risk factor, except for cigarette smoking. Avenues for improving these results are discussed.

Keywords: Machine Learning and Data Mining, Feature Reduction and Manifold Learning, Statistical, Syntactic and Structural Pattern Recognition
Abstract: We propose a nonparametric Bayesian, linear Poisson gamma model for count data and use it for dictionary learning. A key property of this model is that it captures the parts-based representation similar to nonnegative matrix factorization. We present an auxiliary variable Gibbs sampler, which turns the intractable inference into a tractable one. Combining this inference procedure with the slice sampler of Indian buffet process, we show that our model can learn the number of factors automatically. Using synthetic and real world datasets, we show that the proposed model outperforms other state-of-the-art nonparametric factor models.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining, Classification and Clustering
Abstract: We discuss a novel decision tree architecture with soft decisions at the internal nodes where we choose both children with probabilities given by a sigmoid gating function. Our algorithm is incremental where new nodes are added when needed and parameters are learned using gradient-descent. We visualize the soft tree fit on a toy data set and then compare it with the canonical, hard decision tree over ten regression and classification data sets. Our proposed model has significantly higher accuracy using fewer nodes.

Keywords: Pattern Recognition for Bioinformatics
Abstract: To identify microorganisms is of utmost importance in various applications such as medical science and pharmaceutical industry. The technique of Raman spectroscopy is particularly useful in this scenario, since it extracts a high-dimensional molecular fingerprint from samples at hand. Instead of using the complete spectrum, it is often sensible to concentrate on a small number of discriminative dimensions. Apart from providing important molecular insights, this can be beneficial in terms of speed and accuracy. This work studies several state-of-the-art machine learning techniques suitable for feature ranking, many of which have not been used before in the context of Raman spectra classification. Experiments on three different bacteria classification problems show that boosting-based methods and zero-norm support vector machines are especially suited for this challenging task.

Keywords: Pattern Recognition for Bioinformatics, Classification and Clustering
Abstract: Issues to explore in genotypic datasets include the number and characteristic patterns of subpopulations and, possibly, relationships among them. Model-based clustering methods have been adopted to find a number of clusters and the individual assignments. However, they cannot infer genetic relationships among subpopulations the way phylogenetic trees, e.g., the widely-used Neighbor-Joining (NJ) tree, can. In this paper we propose an unsupervised, iterative clustering framework called iNJclust. It performs clustering on an NJ tree with a graph-based partitioning technique. The iterative process enhances the zooming ability and corrects the topology of the final NJ trees. Inference on genetic similarities between subpopulations is also possible. As final outputs, the iNJclust algorithm provides an estimate of the number of clusters, individual assignments, a population tree, as well as sub-trees of the terminal nodes. We illustrate the superior clustering performance of the proposed algorithm using Human 27 populations, bovine 47 breeds, and sheep 28 breeds datasets.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: A standard paradigm to apply graph based representations to computer vision and pattern recognition is to construct a graph from the problem and then formulate the problem in terms of finding cliques in the graph. Many methods have been proposed to extract maximum clique, enumerate all cliques or a number of largest cliques. In this paper we present an approach to a new problem of target clique extraction, i.e., extracting a clique containing a particular vertex. This approach is based on the dominant set clustering and the recently proposed Infection and Immunization Dynamics. We intervene in the game evolution process and gear the convergence towards the target clique. Experiments validate the effectiveness of our approach.

Keywords: Pattern Recognition for Surveillance and Security, Detection, Separation and Segmentation, Image and Video Processing
Abstract: It is sometimes desired to obscure background of a person on a video conference or foreground people in a surveillance video. Background subtraction (or foreground detection) methods can help separate desired from undesired planes, however current methods often have errors � holes in foreground or background � especially after lighting changes. We describe a unified approach to video privacy that capitalizes on the realization that private information is often in the image detail, which have edges, rather than in the uniform intensity regions. So a gradient based method for foreground detection can offer both error tolerance and privacy. We show results of error tolerance to lighting change, and degree of privacy gained by foreground and background privacy filters.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Gesture and Behavior Analysis, Pattern Recognition for Bioinformatics
Abstract: Sparse Representation Classification (SRC) performs well in facial expression recognition (FER). However, SRC based methods costs a lot to train large number of examples. Sparse coding based method will be favorable to tackle the large scale facial expression recognition. K-SVD is state of the art sparse coding method. Unfortunately, K-SVD lacks of discrimination capability for it only focus on the representational power. To cover these problems, we apply discriminative K-SVD (D-KSVD) algorithm on Gabor features for facial expression recognition. Comparing with K-SVD, D-KSVD is more effective for it unifies dictionary and classifiers. We construct comprehensive experiments to verify the proposed algorithm on facial expression database JAFFE. Experimental result indicates that the performance of D-KSVD algorithm on Gabor features is effective than the baselines including SRC and K-SVD algorithms.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Bioinformatics, Image and Video Understanding
Abstract: Matching cells over time has long been the most difficult step in cell tracking. In this paper, we approach this problem by recasting it as a classification problem. We construct a feature set for each cell, and compute a feature difference vector between a cell in the current frame and a cell in a previous frame. Then we determine whether the two cells represent the same cell over time by training decision trees as our binary classifiers. With the output of decision trees, we are able to formulate an assignment problem for our cell association task and solve it using a modified version of the Hungarian algorithm.

Keywords: Biometrics
Abstract: We propose a method for extracting and combining small-scale and large-scale illumination insensitive features for face recognition that can work even in the presence of cast shadows. Although several methods have been proposed to extract such features, they are not designed to handle severe lighting variation on a face and thus fail to work if cast shadows are present. In this paper, we extend quotient image-based illumination normalization by explicitly taking cast shadows into account so that illumination insensitive large-scale features can be obtained. The experimental results demonstrated that the proposed method achieves favorable normalization results under difficult illuminations with cast shadows.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: The evaluation of machine learning algorithms is commonly based on statistical significance tests. However, the suitability of such tests is often questionable. We propose null QQ plots as a simple yet powerful graphical alternative to significance testing. Using ten benchmark data sets, we demonstrate that these plots concisely summarize the essential results from a comparative classification study, while they are easy to produce and interpret.

Keywords: Pattern Recognition for Bioinformatics, Machine Learning and Data Mining, Pattern Recognition for Search, Retrieval and Visualization
Abstract: In this paper, we propose an approach for automated mitosis detection, which provides critical information during performing breast cancer prognosis. Essentially, the problem of mitotic detection involves irregular shape object classification. It is a very challenging task. In this paper, a novel algorithm, named eXclusive Independent Component Analysis (XICA) is proposed. The XICA is an extension of a generic ICA, but focusing the components of differences (called exclusive basis set) between two classes of training patterns rather than the major (independent) components. Based on the residuals obtained from the relative computing involving the exclusive basis set of the relative training patterns, the automated mitosis detection is performed. By computing the residual of the relative exclusive basis set, we are able to classify the given testing patterns. The proposed approach has been tested on a mitosis image set provided by a ICPR2012 contest. It contains 226 mitosis in 35 color images. It achieved accurate rate 100% in training patterns and 83.513% in testing patterns.

Keywords: Classification and Clustering
Abstract: In this paper, we propose a novel algorithm for the problem of predicting change-points. We assume that the causes for change-points can be characterized by the time interval between a change-point and its symptom. Based on this assumption, we first generate weak classifiers for capturing each characteristic, and then build an ensemble classifier with the weak classifiers. Experimental results show our algorithm improves the F-measure by 11% in the best case.

Keywords: Machine Learning and Data Mining, Stereo and Image-Based Modeling, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Classical frameworks for global 1D discrete optimisation: dynamic programming (DP) and belief propagation (BP) -- presume well-posed problems with unique solutions. Ill-posed problems, being the most common in applied pattern recognition and computer vision, are regularised to restore well-posedness. However, typical heuristic regularisation does not guarantee that a set of multiple equivalent solutions is reduced to a single solution. An alternative concurrent propagation (CP) proposed in this paper extends the DP to allow for determining whether the problem is well- or ill-posed and storing implicitly in the latter case the entire set of solutions (e.g. for its structural analysis to improve regularisation). The CP, DP, and BP have similar computational complexity.

Keywords: Gesture and Behavior Analysis, Classification and Clustering, Pattern Recognition for Surveillance and Security
Abstract: Sparse representation based classification (SRC) has been widely used for face recognition (FR). Although SRC algorithm is also adopted in human action recognition, the evaluations of different regular terms have not been given. In this paper, we will discuss and evaluate the role of different regular terms of SRC in human action recognition, after that, we propose human action recognition algorithm based on sparse representation induced by and regulations--- called SR-L12. Experiments on well known KTH action dataset show that SR-L12 is much better than that of nearest neighbor (NN), nearest subspace (NS), full-space (NF), SRC and collaborative representation classification (CRC). Moreover, the proposed method is comparable to most of state-of-the-art algorithms for human action recognition.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing, 2D/3D Object Detection and Recognition
Abstract: We developed a motion blur restoration technique for surface orientation images using a correlation image sensor. This system consists of two components; one is ring-shaped modulation illumination for encoding surface orientation into the amplitude and phase of the reflected light intensity, and the other is the three-phase correlation image sensor (3PCIS) for demodulating the amplitude and phase of reflected light. The object motion is formalized by optical flow constraints. It is solved by weighted integral methods (WIM) developed by us, which is a direct algebraic method. The weighted integral method is suitable for correlation image sensor, because exposure time corresponds to integral over time in WIM and reference signals used by 3PCIS correspond to weighted function in WIM. It is demonstrated by both simulation and experiments that modulation imaging with sinusoids can be used to remove motion blur not only in intensity images, but also in normal vector maps.

Keywords: Image and Video Processing, Physics-Based Vision, Vision for Graphics
Abstract: In this paper we develop a practical method for estimating shape and reflectance using only three polarised images. Using polarised light and retro-reflection settings during image acquisition, we separate the diffuse and specular reflectance components using Blind Source Separation without the accurate knowledge of the polariser angle information. Next, we compare the capacities of five chosen reflectance models, and estimate parameters of appropriate models for the two separated components together with their corresponding zenith angles. Finally, we recover surface shape by using a mixture model to match the two zenith angle estimations. We present experiments to demonstrate the validity of the proposed method for a variety of materials, and we show that the proposed method is capable of accurately estimating both shape and reflectance information.

Keywords: Image and Video Processing, Image and Video Understanding, Human Computer Interaction
Abstract: Facial emotion recognition--the detection of emotion states from video of facial expressions--has applications in video games, medicine, and affective computing. While there have been many advances, an approach has yet to be revealed that performs well on the non-trivial Audio/Visual Emotion Challenge 2011 data set. A majority of approaches still employ single frame classification, or temporally aggregate features. We assert that in unconstrained emotion video, a better classification strategy should model the change in features, versus simply combining them. We compute a derivative of features with histogram differencing and derivative of Gaussians and model the changes with a hidden Markov model. We are the first to incorporate temporal information in terms of derivatives. The efficacy of the approach is tested on the non-trivial AVEC2011 data set and increases classification rates on the data by as much as 13%.

Keywords: Detection, Separation and Segmentation
Abstract: In this paper, a novel spatial-temporal multi-scale method (STMSM) is proposed to solve the problem of detecting multiple moving objects on complex background. Moving objects have multi-scale features both in spatial and temporal domain. The motion salience sub-spaces determine the moving features including position, size and trajectory of each moving object, then the problem of detecting moving objects can be transformed into searching optimal sub-spaces with different scales .This paper proposes a recursive algorithm for estimating motion salience in 3D space and an optimal determinant criterion. These can detect multiple objects at different spatial-temporal scales and extract their features on complex background. The experimental results show this method is effective in detecting multiple moving objects.

Keywords: Remote Sensing, Image and Video Processing
Abstract: This paper proposes a new framework of band selection for hyperspectral images. The algorithm is designed for classification purpose. In this work, different subsets of bands are selected for different class pairs. Without prior knowledge of spectral database, we estimate the spectral characteristic of objects using the labeled and unlabeled samples, benefiting from the concept of semi-supervised learning. Under the assumption of Gaussian mixture model (GMM), the vectors of mean values and covariance matrices for each class are estimated. The separabilities for all pair of classes are thus calculated on each band. The bands with the highest separabilities are then selected. To validate our band selection result, support vector machine (SVM) is employed using a strategy of one against one (OAO). Experiments are carried out on a real data set of hyperspectral image, and the results can validate our algorithm.

Keywords: Image and Video Processing
Abstract: The key frame extraction is designed for obtaining a (very) compressed set of video frames that summarizes the essential content of a video sequence. In this paper, a well-known information theoretic measure, the Jensen-Renyi divergence (JRD), is studied to estimate the frame-by-frame distance between consecutive video images, for segmenting shots/subshots and for choosing key frames. Our new key frame extraction method, which is effective and computationally fast, contributes to a good and quick understanding of a large amount of video data.

Keywords: Detection, Separation and Segmentation, Scene Understanding, Segmentation, Color and Texture
Abstract: This paper proposes a corner and skeleton based method for arbitrarily oriented text detection. By calculating the minimum moment of inertia of each candidate text region, we firstly obtain the orientation and minimum bounding box of each connected component. Secondly, based on the fact that corners are frequent and essential patterns in text regions, we propose a geodesic distance between corner and the skeleton of text regions to measure the effective distance between corner and text. Finally, a geodesic distances weighted corner saturation parameter is given to determine which candidate regions are the true text regions. Experimental results in ICDAR 2003 database show that the proposed method can handle the natural scene text of both horizontal and nonhorizontal orientation.

Keywords: Coding and Compression, Image and Video Processing, Image and Video Understanding
Abstract: In this paper we present a method to recover a spectra representation for reproduction and recognition on multispectral imagery. To do this, we commence by viewing the spectra in the image as a mixture which can be expressed in terms of the sample mean and a set of basis vectors and weights. This treatment leads to an MAP approach where the sample means is given by the centers yielded by the application of the k-means clustering algorithm and the basis vectors are the eigenvectors of the corresponding covariance matrix. We compute the weights making use of a linear programming approach. We illustrate the utility of the method for purposes of skin recognition and spectra reconsruction.

Keywords: Image and Video Understanding
Abstract: Along with the ever-growing Web, horror video sharing through the Internet has interfered with our daily life and affected ours, especially children's, health. Most of current horror video filtering researches pay more attention to the extraction of global features or selection of an optimal classifier, while neglecting the underlying contexts in a scene. In this paper, a novel cost-sensitive sparse coding (CSC) model is proposed to address the context inside scene and interrelations between audio-visual features simultaneously. The model essentially includes two aspects: one is to construct inner contextual structure among frames from same scene based on a graph; the other one is to extend the classic sparse coding technique into a cost-sensitive sparse coding model for graph pattern classification as well as audio-visual features fusion through graph kernel. The experiments on various video scenes demonstrate that our method's performance is superior to the other existing algorithm.

Keywords: Remote Sensing
Abstract: In this paper, we propose a new method for remote sensing image pan-sharpening which is based on weighted red-black (WRB) wavelet and adaptive principal component analysis (PCA), where the adaptive PCA is used to reduce spectral distortions and the utilization of WRB wavelet is used to extract the spatial details in PAN images. To reduce the artifacts and spectral distortions in the pan-sharpened images, which were caused by the local instabilities and dissimilarities in the PAN and MS images, a local process strategy incorporating detail enhancement is introduced. The proposed method is tested on two datasets both acquired by QuickBird and compared with the existing methods. Experimental results show that our method can provide promising fused MS images at a high spatial resolution.

Keywords: Detection, Separation and Segmentation, Segmentation, Color and Texture, 2D/3D Object Detection and Recognition
Abstract: Recently, there are a wide range of cameras used around the world in both indoor and outdoor settings. Although precise color reproducibility is an important requirement, the colors of captured images often seem to be different from those in the original scenes. To evaluate color reproducibility problems, the Macbeth color chart, as shown in Fig. 1, is frequently used. In conventional color correction processes, a color chart is first placed near a target and an image including the chart is captured. Image colors are then corrected based on color deviations of patches on the chart from their true color. However, since the location of the color chart in the image is not readily apparent in such images, color analysis and correction is normally done manually. In this paper, automatic and fast color chart detection method is proposed. In the conventional approach to detecting a color chart, a special chart is used to determine the direction. The difficulty in detecting the chart lies in the fact that its position is unknown and its size within an image is not so large. In addition, in an uncorrected image, the chart colors may deviate from their true values. To detect a color chart precisely and automatically, we propose a method based on a colored pixel spotting approach based on the color array in the chart. Using this array information as a constraint leads to a low computational cost, which makes this method suitable for use in many types of cameras. The effectiveness of this algorithm is confirmed using a 167-image dataset that includes several sizes and rotations of color chart placed at arbitrary locations.

Keywords: Image and Video Processing
Abstract: In this paper we present a novel method for single super-resolution (SR). Given the input low-resolution image, we create a pyramid pair: the ground truth pyramid and the interpolated pyramid. Our method aims to model a relationship between pixel value in the ground truth pyramid and its corresponding 8- neighborhood vector in the interpolated pyramid using Gaussian Mixture Model (GMM). Each pixel in final high-resolution image is predicted by its corresponding 8- neighborhood vector through the trained GMM. Unlike the prior example-based SR method, our algorithm just utilizes the information of input image rather than the external image database. Our proposed algorithm achieves much better results than many state of the art algorithms in terms of both PSNR and visual perception.

Keywords: Detection, Separation and Segmentation
Abstract: To efficiently detect all the possible linear features, a multi scale multi structuring element top-hat transform based algorithm is proposed in this paper. The algorithm is divided into two parts: the multi scale multi structuring element top-hat transform and post-processing. In the multi scale multi structuring element top-hat transform, multi scales of multi structuring elements with increasing sizes are used by the top-hat transform to extract the useful information of linear features. In the post processing, the detected linear feature regions are binarized, firstly. Then, the small noise regions are removed. After that, the final linear feature regions are thinned to form the final binary detected linear features. Experimental results show that, the proposed algorithm could efficiently detect all the possible linear features of different types of images and could be widely used for linear feature detection in different applications.

Keywords: Enhancement, Restoration and Filtering
Abstract: Matting and super-resolution of frames from an image sequence have been studied independently in the literature. We propose a unified formulation to solve both inverse problems by assimilating matting within the super-resolution model. We adopt a multi-frame approach which uses data from adjacent frames to increase the resolution of the matte as well as foreground.

Keywords: Remote Sensing
Abstract: In this paper, a novel example based method is proposed to solve the remote sensing pan-sharpening problem, utilizing an implicit non-parametric learning framework. The high resolution (HR) and downsampled panchromatic (PAN) images are used to train the high/low resolution patch pair dictionaries. Based on the perspective of locally linear embedding (LLE), every patch in each multi-spectral (MS) image band is modeled by its K nearest neighbors in patch set generated from low resolution (LR) PAN image, and this model can be generalized to the HR condition. The intended HR MS patch is reconstructed from the corresponding neighbors in HR PAN patches. Finally, the HR MS images are recovered by stitching these patches together. Two datasets of images acquired by QuickBird satellite are used to test the performance of the proposed method. Experimental results show that the proposed method performs well in preserving spectral information as well as spatial details.

Keywords: Multimedia Analysis, Indexing and Retrieval, Image and Video Processing
Abstract: Analysis of composite shapes recently receives increasing amount of research attention. Different from a silhouette, a composite shape rarely contains a complete envelope. In the paper, we propose a novel envelope extraction algorithm based on the Delaunay triangulation for composite shapes. By analyzing the spatial relationship among individual components of contours and their concavities, we establish new models to describe the envelope edges and their corresponding local enclosed regions. These new models are then used to extract accurate envelopes for composite shapes. We then apply the extracted envelopes to improve shape classification used in shape retrieval. The experimental results show that our algorithm effectively boosts existing shape retrieval algorithms.

Keywords: Image and Video Processing, Enhancement, Restoration and Filtering
Abstract: Sparsity-based super-resolution has attracted lots of attention. Due to the high dimensionality of image data, sparsity-based methods are often in a patch-wise manner and simply impose the smoothness constraints on the overlapped regions between reconstructed patches. However, the imposed smoothness constraint is commonly weak to regularize super-resolution problem when the observed low-resolution image loses structure information. In this paper, we propose to improve the performance of the sparsity-based method by incorporating the structural correlations between neighboring patches. Concretely, the structural information is contained by the dictionary atoms which are used to sparsely represent the image patches. Incorporating the correlations of dictionary atoms into the basic sparse coding, a structural sparse coding algorithm is proposed. Experimental results demonstrate that the proposed algorithm outperforms the sparsity-based baseline in both objective and subjective quality.

Keywords: Image and Video Understanding, Scene Understanding, Cognitive and Embodied Vision
Abstract: This paper proposes a method for estimating the human performance of pedestrian detectability from in-vehicle camera images in order to warn a driver of the positions of pedestrians in an appropriate timing. By introducing features related to visual search and motion of the target, the proposed method estimates the detectability of pedestrians accurately. Support Vector Regression (SVR) is used to estimate the detectability. Here, SVR is trained using features calculated by the proposed method with the ground truth obtained through experiments with human subjects. From experiments using in-vehicle camera images, we confirmed that the proposed features were effective to estimate the detectability of pedestrians.

Keywords: Image and Video Processing, Image and Video Understanding, Multimedia Analysis, Indexing and Retrieval
Abstract: The rapid growth of multimedia data containing rich textual information demands for efficient indexing and retrieval techniques. In this paper, we propose a hybrid approach based on a combination of supervised and unsupervised techniques for the detection of horizontally aligned artificial Urdu text appearing in video images. First, we use an unsupervised approach to detect potential text regions which are later validated by a supervised method. In the first step, edge features followed by morphological operations are used to identify the candidate text regions. These regions are further refined by using edge density and geometrical filters. In the next step, these detected text regions are validated by an Artificial Neural Network which is trained on example text and non-text regions. The effectiveness of the proposed system is evaluated on a dataset of 500 images reading promising results.

Keywords: Enhancement, Restoration and Filtering
Abstract: This paper presents a learning-based method called image super-resolution (SR) for generating a high-resolution (HR) image from a single low-resolution (LR) image. Recent research investigated the image SR problem using sparse coding, which is based on good reconstruction of any image local patch by a sparse linear combination of atoms from an overcomplete dictionary. However, sparse-coding-based SR (ScSR) generally takes a significant amount of computational time to compute an HR image. Further, it can yield only a global dictionary D = [Dh; Dl] by jointly training the concatenated HR and LR image local patches, which results in no accurate correspondence between the HR and LR dictionaries. Therefore, we propose the generation of an HR image using a linear combination of several anchor points (codes) for a local patch based on locality-constrained linear coding (LLC), which is a fast implementation of local coordinate coding (LCC). In the proposed LLC-based strategy, each local patch is represented by a weighted linear combination of its nearer codes in a predefined codebook, and the linear weights become its local coordinate coding. Experimental results show that the recovered HR images with our proposed approach can achieve comparable performance at a processing time much shorter than those of conventional methods.

Keywords: Detection, Separation and Segmentation
Abstract: Feature plays an important role in pedestrian detection, and considerable progress has been made on shape-based descriptors. However, color cues have barely been devoted to detection tasks, seemingly due to the variable appearance of pedestrians. In this paper, Color Maximal-Dissimilarity Pattern (CMDP) is proposed to encode color cues by two core operations, i.e., oriented filtering and max-pooling, which emulate the functions of primary visual cortex (V1). The exten- sively experimental results reveal that the biologically-explainable encoding scheme increases the invariance of color cues, and outperforms the state-of-the-art color descriptor in terms of both accuracy and speed.

Keywords: Multimedia Analysis, Indexing and Retrieval, Motion, Tracking and Video Analysis, Image and Video Processing
Abstract: By segmenting moving objects out and then densely stitching them into background frames, video synopsis provides an efficient way to condense long videos while preserving most activities. Existing video synopsis methods, however, often suffer from either high computation cost due to global energy minimization or unsatisfactory condense rate to avoid loss of important object activities. To address these problems, a tracking based fast online video synopsis approach is proposed in this paper which makes following three main contributions: 1) an online formulation of the video synopsis problem which makes the approach very fast and scalable to endless surveillance videos with reduced chronological disorders, 2) a tracking based schema which can preserve most object activities, and 3) a complete optimization process from both temporal and spatial redundancies of the video which results in much higher condense rate and less object conflict rate. Experimental results demonstrate the effectiveness and efficiency of proposed approach compared to the traditional method on public surveillance videos.

Keywords: Multimedia Analysis, Indexing and Retrieval, Features and Image Descriptors
Abstract: Recent work on Content-Based Image Retrieval (CBIR) have presented alternative methods for fast image indexing and retrieval using Bags of Visual Words (BoVW). In such methods, images are represented as sets of visual words, which can be indexed and searched using well-known text retrieval techniques, allowing fast search on large image databases. In this paper we propose a novel method based on BoVW that improves over current methods by using a new kind of local color descriptor, which we call SDLC, that encodes the most predominant color occurrences in blocks of different im- age regions. We report results of experiments we per- formed with two publicly available image databases. The results indicate that the use of SDLC led to a quite competitive CBIR method in comparison to the state- of-the-art.

Keywords: Multimedia Analysis, Indexing and Retrieval, Pattern Recognition for Search, Retrieval and Visualization, Classification and Clustering
Abstract: 3D-model analysis plays an important role in numerous applications. In this paper, we present an approach for 3D-model retrieval by creating index of closed curves in R3 generated from the center of a 3Dmodel, using a commute time mapping function. Our mapping function respects important properties in order to compute robust closed curves. Each curve describes a small region of the 3D-model. To describe all the mesh, we compute a set of indexed closed curves. These curves lead to creates an invariant descriptor to different transformations. Then we compute the distance between models by comparing the indexed curves. In order to evaluate our method, we used shapes from SHREC 2012 database. The results show the robustness of our method on various classes of 3D-models with different positions.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing
Abstract: In this paper we present a methodology for nonlocal processing of 3D colored point clouds using regularization of functions defined on weighted graphs. To adapt it to nonlocal processing of 3D data, a new definition of patches for 3D point clouds is introduced and used for nonlocal filtering of 3D data such as colored point~clouds. Results illustrate the benefits of our nonlocal approach to filter noisy 3D colored point clouds (either on spatial or colorimetric information).

Keywords: Speech and Audio Analysis
Abstract: Emotion recognition from Speech has been a very active research topic in pattern recognition. In this paper, we investigate the use of kernel reduced-rank regression (KRRR) model to address the emotion recognition problem from speech. KRRR is a nonlinear extension of the linear reduced-rank regression (RRR) model via the kernel trick, in which a kernel mapping is used for the multivariable of RRR. To find the optimal kernel for KRRR, a kernel optimization algorithm is also proposed in the paper. To evaluate the performance of the proposed method, we conduct extensive experiments on the Berlin emotional database. The experimental results confirm the effectiveness of the proposed method.

Keywords: Multimedia Analysis, Indexing and Retrieval, Image and Video Understanding, Image and Video Processing
Abstract: Video highlight recognition is the procedure in which a long video sequence is summarized into a shorter video clip that depicts the most "salient" parts of the sequence. It is an important technique for content delivery systems and search systems which create multimedia content tailored to their users' needs. This paper deals specifically with capturing highlights inherent to sports videos, especially for American football. Our proposed system exploits the multimodal nature of sports videos (i.e. visual, audio, and text cues) to detect the most important segments among them. The optimal combination of these cues is learned in a data-driven fashion using user preferences (expert input) as ground truth. Unlike most highlight recognition systems in the literature that define a highlight to be salient only in its own right (globally salient), we also consider the context of each video segment w.r.t. the video sequence it belongs to (locally salient). To validate our method, we compile a large dataset of broadcast American football videos, acquire their ground truth highlights, and evaluate the performance of our learning approach.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing, Pattern Recognition for Art, Cultural Heritage and Entertainment
Abstract: In this paper, a novel histogram-based model for contrast enhancement is proposed. Based on our analysis about the relationships of histogram with contrast, we establish a model which 1) achieves contrast enhancement by an optimal transform of histogram, 2) gives two metrics called contrast gain and nonlinearity of transform to measure the strength of enhancement and the seriousness of distortion caused by enhancement respectively. The ratio of the two proposed metrics not only gives a guidance for the configuration of parameter in the algorithm, but also provides a useful measurement for contrast distortion, which can be a potential solution to judge whether the contrast of an image is optimal. Experimental results show the superior performances of the proposed algorithm in image enhancement.

Keywords: Image and Video Understanding, Image and Video Processing, Detection, Separation and Segmentation
Abstract: In this work we present a novel approach which combines semantic information with low level features extracted from a complex video scene. The proposed method for video scene understanding relies on a bag-of-words approach, in which, typically, visual words contain information of local motion, but information regarding what generated such motion is discarded. Instead, in our framework, the semantic information is embedded in the visual words and it allows to automatically obtain semantic categorization of the scene. We show the effectiveness of our method in a traffic analysis scenario: in this case, two main semantic classes, pedestrians and vehicles, are discovered.

Keywords: Detection, Separation and Segmentation, Image and Video Processing
Abstract: In this paper, we proposed a novel active contour method for image segmentation, which utilizes the advantages of the GAC and the LRAC methods. We consider the smoothing force of the GAC method and local region-based force of the LRAC method. The advantages of our method are as follows. First the proposed method a new region-based signed pressure force function, which can efficiently stop the contours at weak boundary. Second the proposed method can be handle the heterogeneous texture objects and able to reach into deep concave shapes. Finally, the proposed formulation can be easily implemented by simple finite difference scheme and is computationally more efficient and accurate. The proposed method has been applied to both synthetic and real images.

Keywords: Detection, Separation and Segmentation, Speech and Audio Analysis, Enhancement, Restoration and Filtering
Abstract: We present a sparse representation-based method for detecting adventitious lung sounds in low-quality auscultation signals. Since the noise cannot be represented sparsely by any bases, we can extract clear breath sounds and adventitious sounds from noisy electronic auscultation signals via the sparse representation. Using these clear sound components, we measure the level of abnormality, and robustly detect adventitious sounds with pulsating waveforms, a.k.a crackles. We have experimentally confirmed that our detection achieves an average precision of about 85 percents regardless of nose level.

Keywords: Remote Sensing, Classification and Clustering, Machine Learning and Data Mining
Abstract: We present a new method for the spectral-spatial classification of hyperspectral images, by means of morphological features and manifold learning. In particular, mathematical morphology has proved to be an invaluable tool for the description of remote sensing images. However, its application to hyperspectral data is problematic, due to the absence of a complete lattice structure at higher dimensions. We address this issue by following up previous experimental indications on the interest of classwise orderings. The practical interest of the proposed approach is shown through comparison on the Pavia dataset with Extended Morphological Profiles, against which it achieves superior results.

Keywords: Multimedia Analysis, Indexing and Retrieval, Image and Video Processing, Image and Video Understanding
Abstract: In this paper, a method named histogram intersection metric learning from scene tracks is proposed for automatic organizing people in videos. We make the following contributions: (i) learning histogram intersection distance instead of Mahalanobis distance for widely used face features; (ii) learning the metric from scene tracks without manually labeling any examples, which enables learning across large variations in pose, expression, occlusion and illumination with small number of face pairs and can distinguish different people powerfully. We firstly test face identification, track clustering, and people organization on a long film, then individual retrieval based on people organization from a large video dataset is evaluated, demonstrating significantly increased search quality with respect to previous approaches.

Keywords: Detection, Separation and Segmentation, Speech and Audio Analysis, Speech and Audio Processing
Abstract: A medical staff needs to check sputum accumulation in patient's respiratory tract by lung sounds auscultation at any time, and it is the big burden for the staff. This paper aims to develop a system which notifies appropriate timing for the tracheal suction for the medical staff by analyzing lung sounds of the patients. We present a novel framework about automatic sputum detection from lung sounds. We proposed the sparse representation of audio features to realize robust detection in real environment. We showed the effectiveness of our proposed method for three patients in an ICU of Gifu University Hospital, where the recorded lung sounds included electronic beeps, human voices, and other various noises.

Keywords: Detection, Separation and Segmentation, Low-Level Vision, Features and Image Descriptors
Abstract: This paper proposes a robust detection method for circular objects in noisy and inhomogeneous contrast image. This method detects circular objects not by the difference in image intensities between the object interior and its surrounding, but by the separability and uniformity of the image intensity distributions as calculated by Bhattacharyya Coefficient. The proposed method can detect obscure and textured circular objects, both of which are challenges for conventional methods. In addition, this method does not incur the cost of texture learning. Experiments demonstrate the effectiveness and robustness of the proposed method.

Keywords: Image and Video Processing
Abstract: In this paper, a computation-efficient adaptive support-window scheme is proposed to approximate the conventional bilateral filtering. The difference is that the pixel-wise weights in bilateral filter are thresholded to be only 0 or 1. This results in an adaptive support window, depending on the local image structure of the anchor pixel. A cross-based algorithm is devised to achieve adaptive support window. Experiments show that both noise removal and edge-preserving can be also achieved using our proposed filter. By computing integral images during data aggregation, our algorithm is capable of achieving constant-time complexity regardless of the shape of the support window. Experiments demonstrate that our proposed computing scheme can reduce up to 98% of execution time with respect to the traditional bilateral filter.

Keywords: Multimedia Analysis, Indexing and Retrieval, Image and Video Understanding, Pattern Recognition for Search, Retrieval and Visualization
Abstract: In this paper, we propose a method to recognize food images which include multiple food items considering co-occurrence statistics of food items. The proposed method employs a manifold ranking method which has been applied to image retrieval successfully in the literature. In the experiments, we prepared co-occurrence matrices of 100 food items using various kinds of data sources including Web texts, Web food blogs and our own food database, and evaluated the final results obtained by applying manifold ranking. As results, it has been proved that co-occurrence statistics obtained from a food photo database is very helpful to improve the classification rate within the top ten candidates.

Keywords: Image and Video Processing, Detection, Separation and Segmentation, Enhancement, Restoration and Filtering
Abstract: A computational saliency model utilizing bio-inspired features for spatiotemporal saliency is presented in this paper. We first propose distributed opponent oriented energy for compact local dynamic texture description motivated by Human Vision System. Then, we integrate the derived motion characterization and a revised self-resemblance saliency framework. High effectiveness and efficiency of the proposed method is extensively demonstrated both qualitatively and quantitatively, for background subtraction in the cases of extremely dynamic scenes and camera jitter. In terms of the trade-off between accuracy and computation cost, our method achieves competitive results in contrast to the state-of-art algorithm.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing
Abstract: The distribution of wavelet modulus maxima across wavelet scales can be used to characterize edges in an image. In this paper, we present a novel algorithm that performs image enhancement by mapping the distribution of the wavelet modulus maxima of the blurred image to that of a generic sharp image. Experimental results confirm that the proposed algorithm is able to perform image enhancement without introducing unpleasant visual artifacts.

Keywords: Enhancement, Restoration and Filtering, 2D/3D Object Detection and Recognition, Pattern Recognition for Search, Retrieval and Visualization
Abstract: We propose a new Semi-Lagrangian scheme for the area-preserving Mean Curvature flow. The model uses a level set framework to propagate a closed planar curve. The corresponding flow has been proposed by Sapiro & Tannenbaum cite{ST}. The scheme has the advantage to allow large time step still maintaining a good accuracy. We apply the algorithm to recover the original shape of a synthetic image to which artificial noise has been added. The numerical result shows that the area is preserved during the filtering process.

Keywords: Image and Video Processing, Enhancement, Restoration and Filtering, Low-Level Vision
Abstract: A general method for image contrast enhancement and noise reduction is proposed in this paper. The method is developed especially for enhancing images acquired under very low light conditions where the features of images are nearly invisible and the noise is serious. By applying an improved and effective image de-haze algorithm to the inverted input image, the intensity can be amplified so that the dark areas become bright and the contrast get enhanced. Then, the joint-bilateral filter with the original green component as the edge image is introduced to suppress the noise. Experimental results validate the performance of the proposed approach.

Keywords: Image and Video Processing, Enhancement, Restoration and Filtering
Abstract: In this study, a visual attention region determination approach for H.264 videos using spatiotemporal features is proposed. After Gaussian filtering in Lab color space, the phase spectrum of Fourier transform (PFT) is used to generate the spatial saliency map of each video frame. On the other hand, the motion vector fields from an H.264 video bitstream are backward accumulated and the phase spectrum of Fourier transform (PFT) is used to obtain the temporal saliency map of each video frame. Then, the spatial and temporal saliency maps of each video frame are combined to obtain its spatiotemporal saliency map using adaptive fusion. Finally, a visual attention region determination scheme is used to determine visual attention regions (VARs) of each video frame. Based on the experimental results obtained in this study, the performance of the proposed approach is better than that of two comparison approaches.

Keywords: Remote Sensing
Abstract: We present a method for solving the previously unstudied problem of localizing a set of receivers and directions from transmitters placed far from the receivers, measuring unsynchronized time of arrival data. The same problem is present in node localization of microphone and antenna arrays. The solution algorithm using 5 receivers and 9 transmitters is extended to the overdetermined case in a straightforward manner. Degenerate cases are shown to be when i) the measurement matrix has rank 4 or less or ii) the directions from the transmitters to the receivers lie on an intersection between the unit sphere and another quadric surface. In simulated experiments we explore how sensitive the solution is with respect to different degrees of far field approximations of the transmitters and with respect to noise in the data. Using real data we get a reconstruction of the receivers with a relative error of 14%.

Keywords: Image and Video Processing, Scene Understanding, Pattern Recognition for Surveillance and Security
Abstract: Video summarization, which has a tremendous usage area that spreads from information retrieval to data compression, plays a crucial role in the multimedia understanding. In recent years, with the explosion of the number of videos and their area of use, video summarization became a must to signify. Therefore, this work introduces a novel approach for the summarization problem which is based on human movement understanding. Proposed system presents efficient video knowledge extraction, especially for surveillance cases. Human centric videos are analyzed with histogram of oriented gradients as feature extractor and optical flow as motion descriptor. Above these, a template matching algorithm implemented in a shrinkable and stretchable manner to search for periodicity and thereby extract patterns. Summarization is built up on the validation of these extracted patterns with a correlation based search-through subsystem.

Keywords: Multimedia Analysis, Indexing and Retrieval, Detection, Separation and Segmentation, 2D/3D Object Detection and Recognition
Abstract: This paper presents a probabilistic approach for logo detection and localization in natural scene images. Two probability distributions are computed, one considering the features extracted from the key points located inside a region and the second refers to the shape geometry defined by the key points. The barycentric co-ordinates are considered to define the shape statistics. The performance of the proposed approach has been reported on two publicly available datasets. Logo detection is tested on BelgaLogos and shown that statistically significant improvement is achieved over two recently proposed methods. Logo localization efficiency has been tested on Flickr Logos 27.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing
Abstract: Depth maps captured by Kinect-like cameras are lack of depth data in some areas and suffer from heavy noise. These defects have negative impacts on practical applications. In order to enhance the depth maps, this paper proposes a new inpainting algorithm that extends the original fast marching method (FMM) to reconstruct unknown regions. The extended FMM incorporates an aligned color image as the guidance for inpainting. An edge-preserving guided filter is further applied for noise reduction. To validate our algorithm and compare it with other existing methods, we perform experiments on both the Kinect data and the Middlebury dataset which, respectively, provide qualitative and quantitative results. The results show that our method is efficient and superior to others.

Keywords:
Abstract: In this talk, I first focus on the needs for development of novel techniques for patient and process specific intra-operative imaging and visualization and present some of our latest results as exemplary cases. As novel intra-operative and multi-modality imaging techniques provide the surgical crew with rich co-registered information, their appropriate visualization and their integration into surgical workflow, their validation and finally their full deployment are becoming active subjects of research in our community. Pattern recognition, computer vision and machine learning techniques are further developed to help recovering and modeling surgical procedures and providing innovative solutions. I will in particular trace the Freehand SPECT and Camera Augmented Mobile C-arm (CAMC) from the early development of research ideas within our multi-disciplinary research laboratories to their deployment in different surgical suites. I will finally show how the 'real world laboratories' at our university hospitals demonstrate their efficiency through the smooth path they pave for bringing advance imaging and visualization techniques into the surgical theatres.

Keywords: Computer-Aided Diagnosis and Surgery, Medical Image Analysis and Registration
Abstract: With the increase in age and diabetes-related eye diseases, there is a rising demand for systems which can efficiently screen and locate abnormalities in retinal images. In this paper, we propose a framework that utilizes a variant of the Maximally Stable Extremal Region method, termed C-MSER, to systematically detect various retinopathy pathologies such as microaneurysms, haemorrhages, hard exudates and soft exudates. Experiments on three real-world datasets show that C-MSER is effective for online screening of diabetic retinopathy.

Keywords: Medical Image Analysis and Registration, Computer-Aided Diagnosis and Surgery
Abstract: Localization of the macula centre is an important step in retinal image analysis, in particular for macular disease. We propose the use of a superpixel-based approach for macular localization. Features are extracted from the superpixels, including a proposed feature which aims to describe the extent of the local region due to the superpixel influence. These features are used to calculate probability estimates to determine the macula centre. We evaluated our results on a large dataset of 728 images comprising of normal, glaucoma and AMD eyes. The results are promising. Our method achieved an average error of 30pixels, with all the detected macula centres within 1/8 disc diameters of the reference ground truth, which is lower than the other methods tested.

Keywords: Medical Image Analysis and Registration, Computer-Aided Diagnosis and Surgery
Abstract: Quantitative analysis of optical coherence tomography volumes is an important tool for both clinicians and researchers. Until now, most work has focused on segmentation of the intraretinal cell layers, but the segmentation of pathological datasets remains challenging. We propose the application of random forest to detect the locations of drusen in the retinal pigment epithelium. This is an important step for further analysis of optical coherence tomography data, for segmentation or otherwise. The presented combination of Bruch�s Membrane segmentation with subsequent sampling around the retinal pigment epithelium is a way to quickly compute discriminative features for classification. The proposed method achieves an AUC of 0.94 on our test set, while keeping the computational complexity at a minimum. This makes a clinical setup feasible and provides a tool for clinicians and researchers to quantitatively measure disease progression.

Keywords: Vision for Graphics, Physics-Based Vision, Low-Level Vision
Abstract: In this paper, a novel approach to single image super-resolution based on the multikernel regression is presented. This approach aims to learn the map between the space of high-resolution image patches and the space of blurred high-resolution image patches which are the interpolation results generated from the corresponding low-resolution images. Kernel regression based super-resolution approach is promising, but the kernel selection is a critical problem. In order to avoid selecting the kernel via large amounts of crossverifications, the multikernel regression is applied to learn the map function. This approach is efficient and the experimental results show that it manifests a highquality performance in comparison with other superresolution methods.

Keywords: Segmentation, Color and Texture, Features and Image Descriptors, Mixed and Augmented Reality
Abstract: Environment illumination is a key to achieving a realistic visualization of material appearance. One way to achieve such an illumination is an approximation by rendering of the material surface lit by a finite set of point light sources. In this paper we employed visual psychophysics to identify a minimal number of point light sources approximating realistic illumination. Furthermore, we analyzed stimuli images and correlation of their statistics with obtained psychophysical data. Finally, image statistics were identified which can predict such a minimal environment representation for three tested materials, depending on the visual properties of the illumination environment.

Keywords: Vision for Graphics, 2D/3D Object Detection and Recognition, Medical Image Analysis and Registration
Abstract: In this paper, we present a new three-dimensional (3D) tangent estimator by extending the two-dimensional (2D) lambda-maximal segment tangent (lambda- MST) estimator, which has very good theoretical and practical behaviors. We show that our proposed estimator keeps the same time complexity, accuracy and experimental asymptotic behaviors as the original 2D one.

Keywords: Scene Understanding
Abstract: This paper proposes a method for estimating the illuminant spectral power distributions and their positional relationship of multiple light sources under a complex illumination environment. A multiband camera system is used for capturing spectral images of dielectric objects in a scene. Specular highlights are used as a clue for estimating the light source information, which are detected on curved object surfaces with the different object colors. The illuminant spectra of light sources are estimated from the camera data for each highlight areas. Then, the illuminant spectral estimates are obtained for a different set of light sources. Next, positional relationship among the light sources is predicted by classifying the detected highlights and the estimated spectra using the probabilistic relaxation labeling. The feasibility of the proposed method is examined in experiments on real scenes.

Keywords: Segmentation, Color and Texture, 2D/3D Object Detection and Recognition, Vision for Robotics
Abstract: We propose a novel oversegmentation technique for RGB-D images. The visible surface of the 3D geometry is partitioned into uniformly distributed and equally sized planar patches. This results in a classic oversegmentation of pixels into depth-adaptive superpixels which correctly reflect deformation through perspective projection. The advantages of depth-adaptive superpixels are demonstrated by using spectral graph theory to create image segmentations in near realtime. Our algorithms outperform state-of-the-art oversegmentation and image segmentation algorithms both in quality and runtime.

Keywords: 2D/3D Object Detection and Recognition, Classification and Clustering, Feature Reduction and Manifold Learning
Abstract: We present a novel method for shape analysis which represents shapes as probability density functions and then uses the intrinsic geometry of this space to match similar shapes. In our approach, shape densities are estimated by representing the square-root of the density in a wavelet basis. Under this model, each density (of a corresponding shape) is then mapped to a point on a unit hypersphere. For each category of shapes, we find the intrinsic Karcher mean of the class on the hypersphere of shape densities, and use the minimum spherical distance between a query shape and the means to classify shapes. Our method is adaptable to a variety of applications, does not require burdensome preprocessing like extracting closed curves, and experimental results demonstrate it to be competitive with contemporary shape matching algorithms.

Keywords: 2D/3D Object Detection and Recognition, Classification and Clustering, Image and Video Processing
Abstract: Domain adaptation algorithms that handle shifts in the distribution between training and testing data are receiving much attention in computer vision. Recently, a Grassmann manifold-based domain adaptation algorithm that models the domain shift using intermediate subspaces along the geodesic connecting the source and target domains was presented by Gopalan et al. We build upon this work and propose replacing the step of concatenating feature projections on a very few sampled intermediate subspaces by directly integrating the distance between feature projections along the geodesic. The proposed approach considers all the intermediate subspaces along the geodesic. Thus, it is a more principled way of quantifying the cross-domain distance. We present the results of experiments on two standard datasets and show that the proposed algorithm yields favorable performance over previous approaches.

Keywords: Segmentation, Color and Texture, Scene Understanding, Vision for Robotics
Abstract: We present a new method for the detection of multiple homographies in image pairs. Our aim is to show that we can approach the optimal solution in a short time using an approach based on the wellknown RANSAC algorithm. Given feature correspondences between two similar images, our algorithm iteratively generates homography hypotheses using a suitable sampling, optimizes the promising hypotheses and combine them to keep the best models. The performance of the method is demonstrated on synthetic and real data and we show that it outperforms the J-linkage technique for the multiple homographies estimation problem.

Keywords: Stereo and Image-Based Modeling, Motion, Tracking and Video Analysis, Physics-Based Vision
Abstract: We present a method for recovering fast and robustly the 3D shape of inextensible and smooth surfaces from a monocular image. We propose a weighted iterative least squares approach to minimize the reprojection error between 2D-3D point correspondences preserving the 3D lengths. In addition, a local 3D smoothness constraint for each mesh vertex is proposed to increase the robustness to noisy correspondences and occluded or poorly represented facets. Moreover, the proposed method updates automatically the relevance of each constraint in order to maximize the smoothness and minimize the reprojection error. Experimental results shown that our approach obtains accurate results and is faster than state-of-the-art algorithms using similar constraints.

Keywords: Stereo and Image-Based Modeling
Abstract: Photometric stereo algorithms produce a map of normal directions from the input images. The 3D surface can be reconstructed from this normal map. Existing surface reconstruction works often assume the normal map is integrable but contaminated by small scale non-integrable noise. However, real surfaces often contain large discontinuities such as occlusion boundaries and sharp depth changes, which break the integrable assumption commonly made in many works. Here, we propose a method to detect these discontinuities by combining multiple geometric cues with trained classifiers and a simple graph optimization. The surface is then reconstructed with the guidance of these detected discontinuities. Experiments show our method outperforms existing works.

Keywords: Motion, Tracking and Video Analysis
Abstract: In this paper, we present a method of robust tracking by accounting for hard negatives (i.e., distractors) of the tracking target explicitly. Our method extends the recently proposed Tracking-Learning-Detection (TLD) approach [5] in two aspects: (i) When learning the on-line fern detector, instead of using a set of features which are first randomly generated and then fixed throughout the tracking, we utilize a feature selection stage which constantly improves the performance of the detector, especially in tracking articulated objects (e.g.,pedestrians); (ii) To address the diversity of distractors, instead of tracking a target against the whole set of collected negative examples, we account for the hard negatives explicitly, so that tracking drifts are largely prevented when multiple resembled targets appear in videos (e.g., people with white skirts and jeans). Experiments on a series of diverse videos show that our method outperforms TLD.

Keywords: Motion, Tracking and Video Analysis
Abstract: Geometric reconstruction from image collections is a classical computer vision problem. The problem essentially consists of two steps; First, the identification of matches and assembling of point tracks, and second, multiple view geometry computations. In this paper we address the problem of constructing point tracks using graph theoretical algorithms. From standard descriptor matches between all pairs of images we construct a graph representing all image points and all possible matches. Using Gomory-Hu trees we make cuts in the graph to construct the individual point tracks. We present both theoretical and experimental results (on real datasets) that clearly demonstrates the benefits of using our approach.

Keywords: Motion, Tracking and Video Analysis
Abstract: In this paper, we propose an adaptive particle filter method based on the cross-bin matching, which makes use of the fast and robust earth mover�s distance with a new ground distance (EMD) as the similarity measure, for robust visual tracking. In contrast to the traditional bin-by-bin metrics, the cross-bin metric used in the (EMD) is capable of efficiently capturing the intrinsic affinity relationships among samples, resulting in more accurate and effective tracking results. Experimental results demonstrate the effectiveness and robustness of the proposed tracking method in coping with the challenging situations, such as background clutters, occlusions, abrupt motions and jumps.

Keywords: Motion, Tracking and Video Analysis, Classification and Clustering
Abstract: It is suggested how a Markov random field can be used for object tracking with context information. The tracking is formulated as a two layer process. In the first phase, the image is represented by a set of feature points which are tracked by a standard tracker. In the second phase, the proposed semi-supervised learning and labeling algorithm is used to label the points to three classes -- object, background and companion. The object state (pose) is defined by the set of points labeled as the object. The companion represents the object context and contains non-object points with a motion similar to the motion of the object. As initialization, labels of the object points only are provided by a user in the very first frame. Appearance and motion models of the three classes and the labels of the remaining points in the whole video sequence are estimated in a GrabCut fashion. We show, that use of the companion class together with a 3D (space-time) Markov random field helps to identify object points behind full occlusions or under strong appearance changes.

Keywords: Motion, Tracking and Video Analysis, Geometric and Photometric Registration, Vision for Robotics
Abstract: Concentric circles (C2Tag�s) are planar markers which offer great advantages for detection and tracking. As the circular point-pair (CPP) is the geometric information encoded by a C2Tag, this work is focused on factorization techniques for Structure-and-Motion from multiple CPP images. Gathering all of them in a measurement matrix, two issues are addressed: how to scale the existing entries and how to fill the missing ones. We describe a complete algorithm and prove it. We validate our contributions on simulated and real images.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: In this paper, a unified view of the problem of class-selection with Bayesian classifiers is presented. Selecting a subset of classes instead of singleton allows 1) to reduce the error rate and 2) to propose a reduced set to another classifier or an expert. This second step provides additional information, and therefore increases the quality of the result. The proposed framework, based on the evaluation of the probabilistic equivalence, allows to retrieve the class-selective frameworks that have been proposed in the literature. Several experiments show the effectiveness of this generic proposition.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Detection, Separation and Segmentation
Abstract: In order to estimate multiple structures without prior knowledge of the noise scale, this paper utilizes Jensen-Shannon Divergence (JSD), which is a similarity measurement method, to represent the relations between pairwise data conceptually. This conceptual representation encompasses the geometrical relations between pairwise data as well as the information about whether pairwise data coexist in one model's inlier set or not. Tests on datasets comprised of noisy inlier and a large percentage of outliers demonstrate that the proposed solution can efficiently estimate multiple models without prior information. Superior performance in terms of synthetic experiments and pragmatic tests is also demonstrated to validate the proposed approach.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: Multi-label learning is increasingly required by many domains such as text categorization and scene classification. Learning vector quantization (LVQ) offers a simple, power and scalable algorithm for the single-label learning. In this work, we adapt LVQ to solve the multi-label problems called ML-LVQ. It once adjusts two prototypes for each label of the example to minimize the ranking loss approximately for improving the ranking measures. Moreover, we arm with the single-label AdaBoost.MH as the meta-labeler to predict the number of the labels for the test examples, which will benefit the bipartitions measures. Our empirical study on 6 public multi-label benchmark datasets shows that our proposed algorithm ML-LVQ is statistically significantly better than multi-label AdaBoost.MH and multi-label AdaBoost with the single-label AdaBoost.MH as the meta-labeler especially under the evaluations of the one-error and the mac-F1 (p = 0.03).

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: Due to the absence of nodes and edges correspondence between graphs, the existing central clustering algorithms usually perform graph clustering in some embedded spaces, or confine the cluster centers to the set median graphs. In this paper, a centroid-based algorithm is proposed for directly clustering on the graphs, based on a newly defined dissimilarity measure via structure matching. The resulting node correspondences are used to rearrange the graph structures, such that the cluster centroids, i.e., the generalized median graphs, can be defined directly on the graphs and subsequently be optimized in a k-means type process. The experimental results on four real-world graph datasets demonstrate that the new algorithm significantly improves the clustering accuracy, and is able to discover the meaningful generalized median graphs.

Keywords: Classification and Clustering, Machine Learning and Data Mining
Abstract: Many object recognition or concept identification tasks require accurate detection of large number of classes. These applications present enormous chal- lenges to traditional classification methods, which are mostly designed for solving problems with small num- ber of classes. In this paper, we develop a method called recursive non-negative matrix factorization (RNMF) for building a hierarchical label tree over set of classes. The internal nodes of the tree employ linear classifiers to propagate a data instance to its corresponding leaf node, where one or more one class support vector ma- chine (SVM) classifiers is applied to accurately predict its class. Our experiment results show that the proposed method achieves significant gain in test efficiency and comparable accuracy to some of the more expensive la- bel tree learning methods.

Keywords: Mixed and Augmented Reality, Geometric and Photometric Registration, Features and Image Descriptors
Abstract: In this paper, we discuss how the sensors available in modern smartphones can improve 6-degree-of-freedom (6DOF) localization in wide-area environments. In our research, we focus on phones as a platform for large-scale Augmented Reality (AR) applications. Thus, our aim is to estimate the position and orientation of the device accurately and fast - it is unrealistic to assume that users are willing to wait tenths of seconds before they can interact with the application. We propose supplementing vision methods with sensor readings from the compass and accelerometer available in most modern smartphones. We evaluate this approach on a large-scale reconstruction of the city center of Graz, Austria. Our results show that our approach improves both accuracy and localization time, in comparison to an existing localization approach based solely on vision. We finally conclude our paper with a real-world validation of the approach on an iPhone 4S.

Keywords: Geometric and Photometric Registration, Motion, Tracking and Video Analysis, Mixed and Augmented Reality
Abstract: Model based tracking approaches estimate the pose of the object by minimizing the re-projection error. However, when the object has some ambiguity, for instance, rotation invariance, the 3D pose cannot be correctly estimated. This paper proposes a novel method to allow continuous tracking even when the Degrees of Freedom (DoF) of the target object changes, being able to recover one missing DoF. Pose ambiguity test and recovery of the 3D pose by null space search were added into a general model-based tracking algorithm. Experiments were conducted in a synthetic and in the real world environment to validate the proposed method.

Keywords: Human Computer Interaction, Segmentation, Color and Texture
Abstract: One goal of projector-camera system is let human finger be used like a mouse to click and drag objects in the projected content. It requires segmentation of the human palm and fingers in the image data captured by the camera, which is a challenging task in the presence of the incessant variation of the projected video content and the shadow cast by the palm and fingers. We describe a coarse-to-fine hand segmentation method for projector-camera system. After rough segmentation by contrast saliency detection and mean shift-based discontinuity-preserved smoothing, the refined result is confirmed through confidence evaluation. Extensive experimental results are shown to illustrate the accuracy and efficiency of the approach.

Keywords: Occlusion and Shadow Detection, Physics-Based Vision
Abstract: In this paper, we present a method to detect shadows in outdoor scenes. Here, we note that the shadow areas correspond to the diffuse skylight which arises from the scattering of the sunlight by particles in the atmosphere. This yields a treatment in which shadows in the image can be viewed as a linear combination of scattered light obeying Rayleigh scattering and Mie theory. This treatment allows for the computation of a ratio which permits casting the problem of recovering the shadowed areas in the image into a clustering setting making use of active contours. We illustrate the utility of the method for purposes of detecting shadows in real-world imagery and compare our results against a number of alternatives elsewhere in the literature.

Keywords: Human Computer Interaction, Pattern Recognition for Bioinformatics, Machine Learning and Data Mining
Abstract: This paper proposes a novel active learning method for the classification of motor imagery electroencephalogram (EEG) signals. Specifically, we propose an iterative clustering and support vector-based criterion to select samples of high-confidence to construct a robust training set. The common spatial pattern (CSP)-based features are iteratively clustered till the number of support vectors in the cluster is less than a predefined threshold. A predefined number of samples close to the cluster centers are chosen. When such clusters cannot be found, the samples that are of farthest distances to a group of support vectors of class �0� and �1� are alternately chosen. Experimental results on BCI competition IV dataset IIb show superior performance compared with a baseline method, which is 9% increase in accuracy averaged across subjects and training sizes.

Keywords: Cognitive and Embodied Vision, Enhancement, Restoration and Filtering, Low-Level Vision
Abstract: We propose an image sharpening method that automatically optimizes the perceived sharpness of an image. Image sharpness is defined in terms of the one-dimensional contrast across region boundaries. Regions are automatically extracted for all natural scales present that are themselves identified automatically. Human judgments are collected and used to learn a function that determines the best sharpening parameter values at an image location as a function of certain local image properties. We use the Gaussian mixture model (GMM) to estimate the joint probability density of the preferred sharpening parameters and local image properties. The latter are then adaptively estimated by parametric regression from GMM. Experimental results demonstrate the adaptive nature and superior performance of our approach over the traditional Unsharp Masking method.

Keywords: Computational Photography
Abstract: In recent years, various methods have been proposed for recovering depth blur and motion blur by coding camera optics, such as aperture and exposure. However, these methods are limited to deblurring just a single type of blur, such as depth blur or motion blur. In this paper, we propose a method, which enables us to deblur the depth blur and the motion blur simultaneously by coding image capture both in space and time. The validity and the advantages of the proposed method are shown by real-image experiments and quantitative evaluations using lens simulator.

Keywords: Computational Photography, Vision for Graphics
Abstract: Some computational photography techniques have been proposed to control the focus and illumination of captured images. However, the relationship between the techniques have been unclear because they were developed independently for different purposes. In this research we propose a unified framework to explain the computational photography techniques in the computation of an 8-D reflectance field. Moreover, for an 8-D reflectance field we show that the synthetic aperture, the image-based relighting, and the confocal imaging techniques can be realized using the same measuring device.

Keywords: Computational Photography, Features and Image Descriptors, Image and Video Understanding
Abstract: Photo quality assessment has been a popular research topic. Many previous works achieved high classification rates in photo aesthetics assessment by designing new aesthetic features. However, those hand-crafted features sometimes are not describable, or are very time-consuming and thus not applicable for real-time applications. In this paper, we propose aesthetic features with high efficiency to compute. The experimental results show that our proposed features reach considerable performance. The computation consumption for classifying an image is low so that it is possible to realize online assessment in photo capturing and provide instant feedback to users or fulfill photo rating system on portable devices.

Keywords: Computational Photography, Motion, Tracking and Video Analysis, Low-Level Vision
Abstract: We propose a new video upscaling technique that extends the example-based super-resolution frameworks to multiple-frames. Our method relies on repeating patches that can be observed not only inside a single frame but across the whole video. This allows us to encode image patches with over-complete dictionaries that are constructed in a local spatio-temporal neighborhood around that patch.cut{ To exploit temporal coherence between the frames, we estanblish correpondence between them via optical flow. }We demonstrate the ability of our method to produce high-quality results on real videos with HD resolution, and compare it against state-of-the-art super-resolution techniques.

Keywords: 2D/3D Object Detection and Recognition, Features and Image Descriptors, Stereo and Image-Based Modeling
Abstract: In order to realize model-based 3D object recognition, first, we propose a geometric feature extraction method based on a novel gaze modeling. In the modeling process, local surface models are independently estimated for parts of range data restricted by several gaze domains. Hence, since features are independently extracted from each gaze domain, inconsistent or incorrect features may be obtained. Therefore we introduce a stochastic method that enables us to integrate such features by evaluating the reliability of each gaze model. Next, we propose a shape descriptor, curvature distribution image (CDI), to achieve object recognition by surface matching. It is generated based on the ratios between surface curvatures. In order to discuss the performance of 3D shape description by the kind of used curvatures, we have used 5 kinds of curvatures, such as Gaussian curvature, mean curvature, principal curvatures (maximum, minimum) and shape index. Hence, since an object is represented as a set of descriptive CDIs, the recognition is realized by discovering the one that has the most similar CDIs in the database. The main contribution of this paper is experimental analysis of the performance of CDIs generated by various generation parameters.

Keywords: Motion, Tracking and Video Analysis, Detection, Separation and Segmentation
Abstract: Complementary information, when combined in the right way, is capable of improving clustering and segmentation problems. In this paper, we show how it is possible to enhance motion segmentation accuracy with a very simple and inexpensive combination of complementary information, which comes from the column and row spaces of the same measurement matrix. We test our approach on the Hopkins155 dataset where it outperforms all other state-of-the-art methods.

Keywords: Motion, Tracking and Video Analysis, Stereo and Image-Based Modeling, 2D/3D Object Detection and Recognition
Abstract: This paper addresses the problem of 3D tracking deformable surfaces undergoing non-rigid motion from multi-view video sequences. We propose a method that starts from a set of feature points which have been matched across views and time. A data-driven motion propagation technique makes the motion dense enough to give initial guess to the parameter estimation of the vertices of the 3D model. Finally after rejecting some outliers the 3D mesh model is deformed using the estimated vertex motion. Experimental results demonstrate the proposed method is accurate and robust even at a low frame rate and for medium elastic deformations.

Keywords: Motion, Tracking and Video Analysis, 2D/3D Object Detection and Recognition, Vision for Robotics
Abstract: Progress in LiDAR scanning has led to the availability of large scale LiDAR datasets for urban areas. We use such pre-acquired data to determine the poses of 2D monocular cameras highly accurately in realtime. This is achieved by first correctly aligning key-frames of the multi-modal data using a combination of feature and intensity-based 2D/3D registration methods. The online pose is then determined in realtime by densely sampling and tracking features within the 2D video stream. The 3D coordinates of these features are determined by a fast GPU-based backprojection. The observed 2D/3D feature data is then fused using a recursive bayesian filter in order to exploit temporal coherency. The method is evaluated using ground truth camera trajectories and different filter implementations. The proposed registration and filter framework executes at video-framerate and it is up to 15% more accurate then a registration only solution. Applications are numerous and include, for instance, augmented-reality applications, online georeferentiation or metric online 3D reconstruction from monocular video data.

Keywords: 2D/3D Object Detection and Recognition
Abstract: Finding correspondences between two 3D shapes is common both in computer vision and computer graphics. In this paper, we propose a general framework that shows how to build correspondences by utilizing the isometric property. We show that the problem of finding such correspondences can be reduced to the problem of spectral assignment, which can be solved by finding the principal eigenvector of the pairwise correspondence matrix. The proposed framework consists of four main steps. First, it obtains initial candidate pairs by performing a preliminary matching using local shape features. Second, it constructs a pairwise correspondence matrix using geodesic distance and these initial pairs. Next, the principal eigenvector of the matrix is computed. Finally, the final correspondence is obtained from the maximal elements of the principal eigenvector. In our experiments, we show that the proposed method is robust under a variety of poses. Furthermore, our results show a great improvement over the best related method in the literature.

Keywords: Segmentation, Color and Texture, Features and Image Descriptors
Abstract: Traditional sparse coding has been successfully applied in texture and image classification in the past years. Yet such kind of method neglects the influence of the signs of coding coefficients, which may cause information loss in the sequential max pooling. In this paper, we propose a novel coding strategy for ground-based cloud classification, which is named soft-signed sparse coding. In our method, a constraint on the signs is explicitly added to the objective function of traditional sparse coding model, which can effectively regulate the ratio between the number of positive and negative non-zero coefficients. As a result, the proposed method can not only obtain low reconstruction error but also consider the influence of the signs of coding coefficients. The strategy is verified on two challenging cloud datasets, and the experimental results demonstrate the superior performance of our method compared with previous ones.

Keywords: 2D/3D Object Detection and Recognition, Classification and Clustering, Image and Video Processing
Abstract: Object classification is of vital importance to intelligent traffic surveillance. A big challenge is that shooting view changes in different scenes, which leads to sharp accuracy decrease since training and test samples do not follow the same distribution anymore. On the other hand, manual labeling training samples is time and labor consuming. We propose a feature-based transfer learning framework to gap the divergence of different domain distributions with scarce target view samples. Source view samples, following a different but relevant distribution, could be utilized to learn what a good classifier is like by structure learning. At the same time, small amount of target view samples could make a great contribution to reflect the target distribution. Experimental results indicate that our method outperforms traditional approaches when target samples are too scarce to build a strong classifier.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Surveillance and Security
Abstract: This paper focuses on tracking multiple vehicles in real-world traffic videos which is very challenging due to frequent interactions and occlusions between different vehicles. To address these problems, we fall back on superpixel which recently has received great attention in a wide range of vision problems, e.g. object segmentation, tracking and recognition, for its ability of capturing local appearance characteristics of objects and their spatial relations. As a mid-level feature, however, superpixel itself is unable to carry semantic information which may restricts their use in these problems. To this end, we introduce semantic information into superpixel from an offline trained semantic object detector and successfully deploy it into the multiple vehicle tracking problem. The benefits of semantic superpixel include: 1) it gains better temporal coherency of superpixel; 2) the effectiveness and robustness of occlusion handling are improved; 3) benefited from semantic analysis, false targets and false trajectories are significantly reduced. Experiments show significant accuracy improvements of our approach in comparison with existing tracking methods.

Keywords: Motion, Tracking and Video Analysis, Image and Video Understanding, Multimedia Analysis, Indexing and Retrieval
Abstract: In the paper, we present a method of efficiently summarizing UAV video data. Our approach is based on first detecting and tracking moving objects. Significant amera motion usually present in UAV video data is successfully handled by a robust feature-based frame registration technique. We then devise a saliency-based scoring method to score and rank detected object tracks. Object tracks are then grouped into video segments. The final step is to generate a concise summarization and visualization. Experimental results on the VIRAT UAV dataset show that we can accomplish a data reduction rate in excess of 1000 without significantly missing any activities of interest.

Keywords: Features and Image Descriptors, Segmentation, Color and Texture
Abstract: Feature redundancy and loss of local feature are central problems for image classification. Feature selection decreases the feature redundancy by choosing a subset of features and eliminating those with low prediction. The local feature representation is able to highlight objects in an image, thus, overcoming the drawbacks of global features. This paper presents a new method, called the local kernel for feature selection, which integrates a local kernel of the segmented regions into feature selection to provide improved image classification. This is done by integrating the region-based image distance with the kernel of a Bayesian classifier. The proposed method is tested on two standard image databases and the classification results are higher than the current feature selection and classification methods.

Keywords: Machine Learning and Data Mining, Neural Networks, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Generally, learning the parameters of graphical models by using the maximum likelihood estimation is difficult and requires an approximation. Maximum composite likelihood estimations are statistical approximations of the maximum likelihood estimation and are higher-order generalizations of the maximum pseudo-likelihood estimation. In this paper, we propose a composite likelihood method and investigate its properties. Furthermore, we apply this to restricted Boltzmann machines.

Keywords: Machine Learning and Data Mining
Abstract: Active learning traditionally assumes that the oracle is capable of providing labeling information for each query instance. In reality, the oracle might have no information for some queries and cannot provide accurate label but only answers ``I don't know the label''. We focus on this problem and provide a unified objective function to ensure that each query instance submitted to the oracle is the one mostly needed for labeling and the oracle should also have sufficient knowledge to label. Experimental results on real-world and benchmark data sets demonstrate the effectiveness of the proposed design for supporting active learning using oracles with blind knowledge.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition
Abstract: This paper presents a map matching method based on an ideal Hidden Markov Model (HMM) to find a sequence of roads that corresponds to a given sequence of raw GPS points. Our method is a simplification of the more-complex HMM-based method that maintains its capabilities to cope with the noises and sparsity of the raw GPS data. We test the method with the real-world raw GPS data that is publicly available. Experiments show that despite its simplicity, the proposed method performs sufficiently well under sparse GPS points and sparse road network data.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Sharing of hyper-parameters is often useful for multi-task problems as a means of encoding some notion of task similarity. Here we present a multi-task approach for signal recovery by sharing higher-level hyper-parameters which do not relate directly to the actual content of the signals of interest but only to their statistical characteristics. Our approach leads to a very simple model and algorithm that can be used to simultaneously recover multiple natural images with unrelated content. We investigate the advantages of this approach in relation to state of the art multi-task compressed sensing and we discuss our findings.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition, Segmentation, Color and Texture
Abstract: During the last decade the super-modular Pair-wise Markov Networks (SM-PMN) have become a routinely used model for structured prediction. Their popularity can be attributed to efficient algorithms for the MAP inference. Comparably efficient algorithms for learning their parameters from data have not been available so far. We propose an instance of the Analytic Center Cutting Plane Method (ACCPM) for discriminative learning of the SM-PMN from annotated examples. We empirically evaluate the proposed ACCPM on a problem of learning the SM-PMN for image segmentation. Results obtained on two public datasets show that the proposed ACCPM significantly outperforms the current state-of-the-art algorithm in terms of computational time as well as the accuracy because it can learn models which were not tractable by existing methods.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining, Image and Video Processing
Abstract: Infinite Hidden Markov Random Fields have been proposed for image segmentation as a solution to the problem of automatically determining the number of regions in an image; however, the model does not maintain identity of segmented regions among multiple images. In order to identify segmented regions in images, we developed Hierarchical Dirichlet Process Markov Random Fields. Our model maintains global identification of segmented regions in multiple images by incorporating the idea of hierarchical modeling and automatically determines the number of segmented regions in each image. We show an experimental comparison between the previous model and our proposed model by changing the observation features from RGB value to color histogram features.

Keywords: Classification and Clustering, Pattern Recognition for Bioinformatics
Abstract: Embedded mobile systems for analysis and classification become more and more important in the field of sports and sports science. Small and lightweight sensors in sportswear offer the possibility to monitor the athletes in a realistic environment, e.g. during an outdoor run. During the activity, the sportswear can automatically adapt to the current environment and hence optimizes the performance of the athlete. A major need is a running shoe, which can automatically be adapted to the current ground. In this paper, a classification system was developed, which distinguished between different surfaces and inclinations based on inertial sensors. They were placed on the heel of a running shoe and acquired kinematic data of 21 subjects. For each subject, several rounds of an one hour outdoor run were available and were used for the evaluation of the system. The classification system reached a mean classification rate of more than 80 %.

Keywords: Machine Learning and Data Mining, Multimedia Analysis, Indexing and Retrieval
Abstract: The last few years have seen an exponential increase in the amount of multimedia content that is available online thanks to collaborative-online communities such as Flickr,You Tube etc. As opposed to "pure" social networking services these collaborative-online communities not only allow users to create new social links (e.g. add people to one's friend list) but also allow users to contribute multimedia content and engage in content-driven interactions. A good example of this can be seen in Flickr, in general and Flickr Group in particular where users can comment on or "like" an image contributed by another user. This paper looks at utilizing this within group user-user interaction information, along with image meta-data to discover user communities (user-subgroups) that contribute content around specific topics (subgroup-themes) at specific points in time. A good example of this is a group of users (e.g sports fans) contributing content and interacting with each other only at specific times of the year (e.g close to their favorite sporting event). We demonstrate that our proposed generative model Temporal Block-Link LDA is able to successfully extract such user-subgroups, subgroup-themes and associated temporal patterns from data in an unsupervised manner.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Feature Reduction and Manifold Learning
Abstract: Embedded microcontrollers are employed in an increasing number of applications as a target for the implementation of classification systems. This is true for example for the fields of sports, automotive and medical engineering. However, important challenges arise when implementing classification systems on embedded microcontrollers, which is mainly due to limited hardware resources.

In this paper, we present a solution to the two main challenges, namely obtaining a classification system with low computational complexity and at the same time high classification accuracy. For the first challenge, we propose complexity measures on the mathematical operation and parameter level, because the abstraction level of the commonly used Landau notation is too high in the context of embedded system implementation. For the second challenge, we present a software toolbox that trains different classification systems, compares their classification rate, and finally analyzes the complexity of the trained system. To give an impression of the importance of such complexity measures when dealing with limited hardware resources, we present the example analysis of the popular Pima Indians Diabetes data set, where considerable complexity differences between classification systems were revealed.

Keywords: Classification and Clustering, Pattern Recognition for Surveillance and Security, Machine Learning and Data Mining
Abstract: We propose a novel trajectory clustering algorithm which is suitable for online processing of pedestrian or vehicle trajectories computed with a vision-based tracker. Our approach does not require defining distances between trajectories, and can thus process broken trajectories which are inevitable in most cases when object trackers are applied to real world video footage. Clusters are defined as smooth vector fields on a bounded connected set, and cluster distance is based on pairwise distances between vector sets. The results are illustrated on a trajectory set from the Edinburgh Informatics Forum Pedestrian Dataset, on a trajectory set from a public transport junction, and trajectories from an experimental setup in a corridor.

Keywords: Classification and Clustering, Remote Sensing, Image and Video Understanding
Abstract: The huge increase of hyperspectral data dimensionality and information redundancy has brought high computational cost as well as the over-fitting risk of classification. In this paper, we present an automatic band selection and classification method based on a novel wrapper Multiple Improved particle swarm cooperative optimization and support vector machine model (MIPSO-SVM). The MIPSO-SVM model optimizes both the band subset and SVM kernel parameters simultaneously. In the proposed model, the particle swarm is divided into two sub-swarms. And PSO is improved firstly, by the new update strategy of position and velocity. Then the sub-swarms perform the improved PSO (IPSO) for band selection and classifier parameters optimization independently. Finally, in the process of cooperative evolution, extremal optimization (EO) is incorporated to maintain the diversity of swarms and enhance the space exploration ability of the proposed model. Experimental results demonstrate the effectiveness of the proposed method for band selection and classification of hyperspectral images.

Keywords: Feature Reduction and Manifold Learning, Biometrics, Pattern Recognition for Bioinformatics
Abstract: Due to the fact that many objects in the real world can be naturally represented as tensors, tensor subspace analysis has become a hot research area in pattern recognition and computer vision. However, existing tensor subspace analysis methods cannot provide an intuitionistic nor semantic interpretation for the projection matrices. In this paper, we propose Sparse Tensor Principal Component Analysis (STPCA), which transforms the eigen-decomposition problem to a series of regression problems. Since its projection matrices are sparse, STPCA can also address the occlusion problem. Experiment on Georgia tech database and AR database showed that the proposed method outperforms the Multilinear Principal Component Analysis (MPCA) in terms of accuracy and robustness.

Keywords: Machine Learning and Data Mining, Pattern Recognition for Surveillance and Security, Classification and Clustering
Abstract: Unsupervised anomaly detection techniques are becoming more and more important in a variety of application domains such as network intrusion detection, fraud detection and misuse detection. Today, unsupervised anomaly detection techniques are mainly based on quadratic complexity making it almost impossible to apply them on very large data sets. In this paper, an Expectation-Maximization algorithm is proposed which computes the Local Outlier Factor (LOF) incrementally and up to 80% faster than the standard method. Another advantage of FastLOF is that intermediate results can be used by a system already during computation. Evaluation on real world data sets reveal that FastLOF performs comparable to the best outlier detection algorithms although being significantly faster.

Keywords: Gesture and Behavior Analysis, Motion, Tracking and Video Analysis
Abstract: Understanding fish behavior by extracting normal motion patterns and then identifying abnormal behaviors is important for understanding the effects of environmental change. In the literature, there are many studies on normal/abnormal behavior detection in the areas of human behaviour analysis, traffic surveillance, and nursing home surveillance, etc. However, the literature is very limited in terms of normal/abnormal fish behavior understanding especially when natural habitat applications are considered. In this study, we present a rule based trajectory filtering mechanism to extract normal fish trajectories which potentially helps to increase the accuracy of the abnormal fish behavior detection systems and can be used as a preliminary method especially when the number of abnormal fish behaviors are very small (e.g. 40-50 times smaller) compared to the number of normal fish behaviors and/or when the number of trajectories are huge.

Keywords: Classification and Clustering, Pattern Recognition for Surveillance and Security
Abstract: In this paper, we present a novel semi-supervised strategy for machine fault diagnosis. In the proposed method, we select parzen window as the generative classifier and Gentleboost as the discriminative classifier. Compared with SVM, boosting method has a very interesting property of relative immunity to overfitting. In addition, we propose a novel adaptive parzen window algorithm. It employs variational adaptive parzen window rather than a global optimized and fixed window, therefore, more accurate density estimates can be obtained. In experiments, artificial and machine vibration data are used to compare with other algorithms. Our proposed algorithm achieves stronger robustness and lower classification error rate.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Gesture and Behavior Analysis, Handwriting Recognition
Abstract: This paper proposes a new dynamic time warping (DTW) method, called non-Markovian DTW. In the conventional DTW, the warping function is optimized generally by dynamic programming (DP) subject to some Markovian constraints which restrict the relationship between neighboring time points. In contrast, the non-Markovian DTW can introduce non-Markovian constraints for dealing with the relationship between points with a large time interval. This new and promising ability of DTW is realized by using graph cut as the optimizer of the warping function instead of DP. Specifically, the conventional DTW problem is first converted as an equivalent minimum cut problem on a graph and then edges representing the non-Markovian constraints are added to the graph. An experiment on online character recognition showed the advantage of using non-Markovian constraints during DTW.

Keywords: Classification and Clustering, Document Analysis Systems, Machine Learning and Data Mining
Abstract: Non-negative matrix factorization (NMF) is a well known tool for unsupervised machine learning. It can be viewed as a generalization of the K-means clustering, Expectation Maximization based clustering and aspect modeling by Probabilistic Latent Semantic Analysis (PLSA). Specifically PLSA is related to NMF with KL-divergence objective function. Further it is shown that K-means clustering is a special case of NMF with matrix L2 norm based error function. In this paper our objective is to analyze the relation between K-means clustering and PLSA by examining the KL-divergence function and matrix L2 norm based error function.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, 2D/3D Object Detection and Recognition
Abstract: We present a novel probabilistic model for parsing shapes into several distinguishable parts for accurate shape recognition. This shape parsing is based on robust geometric features that permit high recognition accuracy. Although modelling shapes is an inherently uncertain process, our approach is lenient, in that the desired parse of a shape only needs to be within its k most probable parses. Using this set of shape decompositions, we can boost recognition accuracy even further by determining which parts of a shape are common across most views of objects in the same class.

Keywords: Pattern Recognition for Bioinformatics, Medical Image Analysis and Registration, Machine Learning and Data Mining
Abstract: Breast cancer grading of histological tissue samples by visual inspection is the standard clinical practice for the diagnosis and prognosis of cancer development. An important parameter for tumor prognosis is the number of mitotic cells present in histologically stained breast cancer tissue sections.

We propose a hierarchical learning workflow for automated mitosis detection in breast cancer. From an initial training set a pixel-wise classifier is learned to segment candidate cells, which are then classified into mitotic and non-mitotic cells using object shape and texture features. Our workflow banks on two open source biomedical image analysis software: "ilastik" and "CellCognition" which provide a user user friendly interface to powerful learning algorithms, with the potential of making the pathologist work an easier task.

We evaluate our approach on a dataset of 35 high-resolution histopathological images from 5 different specimen (provided by textit{International Conference for Pattern Recognition} 2012 contest on textit{Mitosis Detection in Breast Cancer Histological Images}). Based on the candidate segmentation our approach achieves an area-under Precision-Recall-curve of 70% on an annotated dataset, with good localization accuracy, little parameter tuning and small user effort. Source code is provided.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: We propose a new semi-supervised relational clustering approach, called Semi-Supervised Fuzzy clustering with Learnable Cluster dependent Kernels (SSFLeCK). The proposed algorithm learns the underlying cluster-dependent dissimilarity measure while finding compact clusters. The learned dissimilarity is based on a Gaussian kernel with cluster dependent scaling parameters. SS-FLeCK uses side-information in the form of a small set of constraints on which instances should or should not reside in the same cluster. The proposed algorithm uses only the pairwise relation between the feature vectors. This makes it applicable when similar objects cannot be represented by a single prototype. Using synthetic and real data sets, we show that SSFLeCK outperforms several other algorithms.

Keywords: Machine Learning and Data Mining, Pattern Recognition for Search, Retrieval and Visualization
Abstract: Among ensemble learning methods, stacking with a meta-level classifier is frequently adopted to fuse the output of multiple base-level classifiers and generate a final score. Labeled data is usually split for base-training and meta-training, so that the meta-level learning is not impacted by over-fitting of base level classifiers on their training data. We propose a novel knowledge-transfer framework that reutilizes the base-training data for learning the meta-level classifier without such negative consequences. By recycling the knowledge obtained during the base-classifier-training stage, we make the most efficient use of all available information and achieve better fusion, thus a better overall performance. With extensive experiments on complicated video event detection, where training data is scarce, we demonstrate the improved performance of our framework over other alternatives.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: This paper proposed an unsupervised learning method to learn speech features based on Dynamic Bayesian Networks (DBNs) that accounts for the spatiotemporal dependences in speech signal. Although deep networks have been successfully applied to unsupervised learning features, the structures of the deep networks are often fixed before learning and they fail to capture temporal representation. In this paper, we propose to construct DBNs for unsupervised learning spatial-temporal features from speech data. The experiment results on Tibetan speech data showed the features learned using proposed DBNs outperforms the state-of-art methods in word recognition accuracy.

Keywords: Machine Learning and Data Mining, Pattern Recognition for Search, Retrieval and Visualization, Classification and Clustering
Abstract: In this paper, we propose a new learning paradigm named multitask multiclass privileged information support vector machines. The starting point of our work is mainly based on the success of multitask multiclass support vector machines which cast multitask multiclass problems as a constrained optimization problem with a quadratic objective function. Learning using privileged information is an advanced learning paradigm integrated with the idea of human teaching in machine learning. This paper mainly extends multitask multiclass support vector machines to privileged information learning strategy. Our approaches can take full advantages of the multitask learning and privileged information. Experimental results show that our approaches obtains very good results for multitask multiclass problems.

Keywords: Machine Learning and Data Mining, Pattern Recognition for Bioinformatics
Abstract: This paper proposes a novel regression method based on distance metric learning for human age estimation. We take age estimation as a problem of distance-based ordinal regression, in which age difference is measured by an efficient distance metric. To reach this goal, we propose to learn such a distance metric that can preserve both the ordinal information of different age groups and the local geometry structure of the target neighborhoods simultaneously. Then, the facial aging trend can be truly discovered by the learned metric. Experimental results on the publicly available FG-NET database are very competitive against the state of the art.

Keywords: Classification and Clustering, Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition
Abstract: There are many popular models available for classification of documents like Naive Bayes Classifier, k-Nearest Neighbors and Support Vector Machine. In all these cases, the representation is based on the Bag of words model. This model doesn't capture the actual semantic meaning of a word in a particular document. Semantics are better captured by proximity of words and their occurrence in the document. We propose a new Bag of Phrases model to capture this discriminative power of phrases for text classification. We present a novel algorithm to extract phrases from the corpus using the well known topic model, Latent Dirichlet Allocation(LDA), and to integrate them in vector space model for classification. Experiments show a better performance of classifiers with the new Bag of Phrases model against related representation models.

Keywords: Biometrics, Classification and Clustering, Machine Learning and Data Mining
Abstract: Video-based biometric systems are becoming feasible thanks to advancement in both algorithms and computation platforms. Such systems have many advantages: improved robustness to spoof attack, performance gain thanks to variance reduction, and increased data quality/resolution, among others. We investigate a discriminative video-based score-level fusion mechanism, which enables an existing biometric system to further harness the riches of temporarily sampled biometric data using a set of distribution descriptors. Our approach shows that higher order moments of the video scores contain discriminative information. To our best knowledge, this is the first time this higher order moment is reported to be effective in the score-level fusion literature. Experimental results based on face and speech unimodal systems, as well as multimodal fusion, show that our proposal can improve the performance over that of the standard fixed rule fusion strategies by as much as 50%.

Keywords: Classification and Clustering, Machine Learning and Data Mining
Abstract: Support vector clustering (SVC) is a nonparametric clustering algorithm inspired by support vector machines. Incremental support vector clustering (ISVC) extends the SVC algorithm to an incremental version for the case of large-scale datasets with the assumption of no outliers. In order to tackle the problem of clustering large-scale noisy datasets, this paper proposes the algorithm termed incremental support vector clustering with outlier detection (OD-ISVC). The proposed algorithm consists of two components, namely, incremental support vector (SV) construction and dynamic bounded support vector (BSV) management. We introduce the concept of BSV-pool, where the check and recycle procedure is designed for updating the temporarily stored BSVs and detecting outliers. The experiments on real and synthetic datasets demonstrate the effectiveness and efficiency of our method.

Keywords: Biometrics
Abstract: A new descriptor called Multi-modal Binary Patterns (MMBP) is proposed for face recognition. It balances well important requirements for real-world applications, including the robustness, discriminative power, and the low computational cost. The proposed algorithm has several desirable properties: 1) it captures information from face image in any direction as it is oriented feature, 2) being a spatial multi-scale structure, the descriptor catches not only local but also more global information about object, 3) it is robust to image transformation like variations of lighting, expressions, and 4) it is computationally efficient. In more detail, to catch information in a given direction, a Local Line Binary Pattern (LLBP) based operator is first applied. The MMBP feature is then built by applying a LBP-based self-similarity operator on the values being calculated by LLBP operators across different directions. A Whitened PCA dimensionality reduction technique is applied to get more a compact and efficient descriptor. Experimental results achieved on the comprehensive FERET data set being comparable to state-of-the-art validates the efficiency of our method.

Keywords: Biometrics
Abstract: In this paper, we present a novel face alignment approach in thermal infrared face recognition. The alignment procedure is based on closest point set matching between sketch faces. Linear combination of positional and local pattern features is embedded in the pointwise distance to solve the local minimum problem for ICP due to edge noise in sketch faces. The comprehensive experiments, including intra-class, inter-class and variable expressions, show the alignment accuracy and face recognition performance results of our algorithm compared to manual labelling, ICP and congealing methods.

Keywords: Biometrics, Pattern Recognition for Surveillance and Security, Pattern Recognition for Search, Retrieval and Visualization
Abstract: Several citizen service databases such as, police, national citizen identity, passport and vehicle registration, store both biographical and biometric information containing huge number of records. Achieving scalability and high accuracy for a 1:N person identification task on these databases is a huge challenge. In this work, we propose to use complementary information present in the biographical data along with biometric information of a user to improve 1:N person identification task for large systems. We show that a likelihood ratio based method for score level fusion of the biometric and biographical classifiers results in high accuracy identification as compared to using only the biometric classifiers or the biographical classifiers

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: We propose a nonparametric Bayesian approach to time series alignment. Time series alignment is a technique often required when we analyze a set of time series in which there exists a typical structural pattern common to all the time series. Such a set of time series is typically obtained by repeated measurements of a biological, chemical or physical process. In time series alignment, we are required to estimate a common shape function, which describes a common structural patter shared among a set of time series, and time transformation functions, each of which represents time shifts involved in individual time series. In this paper, we introduce a generative model for time series data in which the common shape function and the time transformation functions are modeled nonparametrically using Gaussian processes and we develop an effective Markov Chain Monte Carlo algorithm, which realizes a nonparametric Bayesian approach to time series alignment. The effectiveness of our method is demonstrated in an experiment with synthetic data and an experiment with real time series data is also presented.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining, 2D/3D Object Detection and Recognition
Abstract: Shape analysis relies on using a finite number of points on the contour of an object to compare the shapes of objects. These points are called landmarks. Hence, when landmarks are not available for analysis, we must place some appropriately on the contour. In this paper, we describe a new method for placing landmarks well on the contours of objects in the same class. The landmarks located by our method are determined by solving an optimization problem. In experiments using line drawings, we demonstrate that our method places landmarks well on each drawing.

Keywords: Pattern Recognition for Bioinformatics, Human Computer Interaction, Image and Video Understanding
Abstract: In this paper, a novel hierarchical algorithm with multi-feature fusion is proposed for facial expression recognition. In this area, many people have proposed many good results, but few of them made good use of the distribution characteristic of facial expression itself. In the analysis of the feature distribution, we find happiness and surprise are clearly separated from the other expressions. So we aim to distinguish these two expressions in the first layer of our algorithm using Gabor features. In the second layer, we use Gabor and LBP features respectively to classify the other five expressions. And a well designed result fusion of two branches is adopted to improve the accuracy. Experiments results on the Cohn-Kanade database show that our algorithm achieves excellent accuracy. Furthermore, our algorithm also performs well in our hybrid database, in which there are extensive variations of expressions. It demonstrates the good generalization ability of our algorithm.

Keywords: Pattern Recognition for Surveillance and Security, Image and Video Processing, Pattern Recognition for Search, Retrieval and Visualization
Abstract: Appearance of individuals across multiple cameras varies a lot due to illumination and viewpoint changes making person re-identification a challenging problem. In this paper, we describe how to model this appearance variation by using a novel Weighted Brightness Transfer Function (WBTF). In combination with powerful low-level features, we show that WBTF leads to large performance improvements by assigning different weights to different BTFs and combining them accordingly. We have compared our algorithm on two public benhmark datasets: VIPeR and CAVIAR4REID dataset, achieving new state-of-the art performance on both datasets.

Keywords: Feature Reduction and Manifold Learning, Graphics Recognition
Abstract: Finding discriminant features is useful for pattern recognition applications. In this work, geometric matching is combined with linear discriminant analysis (LDA) to learn the importance of the features of symbols, and assign weights to these features accordingly. The features are the line segments of the symbols. We use geometric matching within a symbol spotting system to get information on the matching between the line segments of a query symbol and the line segments of the spotted symbols found by the spotting system (both true and false matches). The matching information is used to compute feature vectors for a query symbol. The vectors represent how well the segments of a query are matched to the segments of the true and false matches. Then, LDA is trained on these vectors to get the weights of the line segments of different query symbols. This feature weighting approach is applied in symbol spotting. Using the query weighted features, the spotting system�s precision improves from an average of 71% to an average of 98%, with a speed up factor of 2.1.

Keywords: Pattern Recognition for Surveillance and Security, Statistical, Syntactic and Structural Pattern Recognition
Abstract: The security of web services is nowadays one of the major concerns for Internet users. Web services may manage confidential information, monetary transactions, or even health-critical systems, such as those employed in public airports or hospitals. A key problem of web services is that they should work as expected even in the presence of malicious inputs. Unfortunately, with the increasing complexity of web services, this task becomes more and more challenging.

In this paper we present SuStorID, a multiple classifier system which is able to model legitimate inputs towards web services, given a sample of web traffic. If anomalous inputs are detected, web services are protected according to a set of anomaly templates. Our experiments, performed on a production environment, highlight that our system can accurately detect web attacks and help security operators to protect their web services against known and unknown attacks.

Keywords: Biometrics
Abstract: Fingerprint matching is one of the most widely used biometrics for personal identification. However, the performance of fingerprint identification system is insufficient for many applications. Lots of methods were proposed to improve system performance by introducing more information into matching process. In this paper, we introduced a new kind of information named prior knowledge and proposed a score-level fusion method with prior knowledge for fingerprint matching. The trend and discrimination of scores are used as prior knowledge with sigmoid function to search the optimal fusion parameters. Experimental results show that the proposed prior knowledge is useful for score fusion and fingerprint matching and the score-level fusion algorithm is effective to improve system performance and comparative to the best ones in FVC2004.

Keywords: Biometrics, Pattern Recognition for Surveillance and Security
Abstract: In this paper, we introduce a torus manifold-based temporal super resolution method for gait recognition from low frame-rate videos with view transitions. Given a low frame-rate gait sequence with view transition from an unknown person, we estimate three unknowns: view, phase, and style. We estimate view by walking trajectory and camera information, phase by dynamic programming using multiview exemplar sequences, and style by bilinear model and linear least squares. Once these parameters are known, we can synthesize a high frame-rate sequence corresponding to that unknown person and can use existing methods for gait recognition. Experiments with OU-ISIR multiview gait dataset demonstrate the effectiveness of the proposed method for frame-rates as low as 1 or 2 fps.

Keywords: Biometrics
Abstract: Traditionally, it was difficult to use vein patterns in evidence images for forensic identification, because they were nearly invisible in color images. We proposed a computational method based on skin optics to uncover vein patterns from color images. However, its performance is dependent on the accuracy of the skin optical model. In this paper, we propose an algorithm based on image mapping to visualize vein patterns. It extracts information from a pair of synchronized color and near infrared (NIR) images, and uses a neural network (NN) to map RGB values to NIR intensities. In addition, an NN weight adjustment scheme is proposed to improve the robustness of the algorithm. The proposed algorithm was examined on a database with 300 pairs of color and NIR images collected from the forearms of 150 subjects. The automatic matching results from the proposed algorithm were better than those from our previous method, and comparable to the results from matching NIR images with NIR images.

Keywords: Classification and Clustering, Feature Reduction and Manifold Learning, Features and Image Descriptors
Abstract: Person-dependent appearance changes tend to increase difficulties in automatic facial expression recognition. Although one can use neutral face images to reduce the personal variations, acquisition of neutral face images may not always be possible in real cases. In order to remove the person-dependent influence from expressive images, we propose a dual subspace nonnegative matrix factorization (DSNMF) to decompose facial images into two parts: identity and expression parts. The identity part should characterize person-dependent variations, while the expression part should characterize person-invariant expression features. Our experimental results show that the proposed method significantly outperforms existing approaches on the CK+ and JAFFE expression databases.

Keywords: Biometrics, Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: In this paper we study the problem of score normalization in biometric verification systems. Specifically, we introduce a new class of normalization techniques, which unlike the commonly used parametric score normalization techniques, such as z- or t-norm, make no assumptions regarding the shape of the underlying score distribution. The proposed class of normalization techniques first estimates the relevant score distribution in an impostor-centric manner using kernel density estimation and then maps the estimated distribution to a common one. Our experimental results obtained on the FRGCv2 face database show that the proposed non-parametric score normalization techniques consistently outperform their parametric counterparts when the target distribution takes a log-normal form and that all assessed techniques, i.e., z-, t-, zt- and tz-norms, improve upon the setting where no score normalization is used.

Keywords: Biometrics
Abstract: Human hand is composed of structures called carpal bones, metacarpal bones and phalanges (which form the fingers). Typically, fingerprint matching is used for personal authentication, with images & features obtained from the "tip" of the fingers, ie. distal phalanges (sections, digits). In this study, we report fingerprint minutiae matching results, with images obtained from proximal and middle phalanges. Experiments conducted on a medium-size database, collected using a commercial low-cost optical (distal) fingerprint sensor without any modification, show that, in applications where distal phalanx images are not usable (e.g. due to missing digits, low quality finger surface due to manual labor), non-distal phalanges may provide an acceptable biometric verification source.

Keywords: Gesture and Behavior Analysis, Pattern Recognition for Surveillance and Security, Classification and Clustering
Abstract: In this paper, we propose an approach for human activity categorizing based on the use of optical flow direction and magnitude features. The main contribution of this paper is the feature representation that mirrors the geometry of the human body and relationships between its moving regions when performing activities. The features are quantified using a quantization algorithm. We analyze the performance of two well-known classifiers: the Na�ve Bayes and the SVM. The results show the effectiveness of our approach.

Keywords: Pattern Recognition for Bioinformatics, Pattern Recognition for Surveillance and Security, Biometrics
Abstract: Human faces undergo considerable amount of variations across ages. This paper proposes an age-invariant face verification method by using a Local Classifier Ensemble Model (LCEM). First, reference points are located based on an extended Active Shape Model and faces are aligned afterwards. Second, a face is grouped into several non-overlapping patches and each group is further divided into several overlapping sub-patches. Local classifiers are then trained for each sub-patch and integrated to build an ensemble classifier under a Semi-Naive Bayesian framework. Finally, the proposed approach is evaluated on the MORPH database and experimental results show that such local classifier ensemble achieves significant and robust performance.

Keywords: Pattern Recognition for Surveillance and Security, Image and Video Processing
Abstract: Pedestrian detection is a key problem in many computer vision applications, especially in surveillance and security systems. To this end, information integration from different imaging modalities, such as thermal infrared and visible spectrum, can significantly improve the detection rate in respect to mono-modal strategies. For this reason, an effective fusion scheme is necessary to combine the information presented by multiple sensors. In this paper, we propose a pedestrian classification method based on the multiple kernel learning framework; standard pixel features (such as spatial derivatives) from both imaging modalities are employed to learn several feature-related basic kernels and a compound kernel is found as an optimized linear combination of basic kernels. Finally the compound kernel is used to train an SVM. Experiments performed on the OTCBVS dataset [1], demonstrate that our recipe definitely outclasses a wide set of literature fusion modalities.

Keywords: Biometrics, Image and Video Processing
Abstract: In this paper we propose a watermarking based two-stage authentication framework to enhance biometric system security. The face feature of one individual is embedded into his/her fingerprint image as credibility token. During authentication, the watermark is first extracted to establish data authenticity. If legitimate, the face pattern can further serve as supplemental trait in biometric authentication progress. A wavelet quantization based watermarking method is proposed to robustly embed information in fingerprints while preserving their discriminating features. Meanwhile, a sparse representation based classifier is adopted to efficiently exploit the identity information within face watermarks. Experimental results which evaluate both watermarking and biometric authentication performance demonstrate the effectiveness of this work.

Keywords: Biometrics, Pattern Recognition for Surveillance and Security
Abstract: Iris recognition in less constrained environments is challenging due to the degraded iris images. This paper proposes a novel method fusing multiple cues for iris recognition in the non-ideal imagery. The covariance matrices are used to represent local iris texture property, which capture the correlation of spatial coordinates, intensities, 1st and 2nd-order partial derivatives. The covariance matrices are symmetric positive definite (SPD) which form a Riemannian space rather than a Euclidean one. In the Log-Euclidean framework, the space of SPD matrices is equipped with a linear space structure so that in the logarithmic domain the Euclidean operations are applicable. This enables us to compute the logarithms of covariance matrices, leading to the Log-Euclidean covariance matrices (LECM), which can be handled in common Euclidean operations. The ordinal measure is further used to represent the order relation of iris texture by comparing LECMs at different positions. We finally perform iris matching based on the Hamming distance in which the noise effects are considered. Experiments on challenging databases show the effectiveness of the proposed method.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition, Classification and Clustering
Abstract: When a multi-label classifier outputs a real-valued score for each class, a well known design strategy consists of tuning the corresponding decision thresholds by optimising the performance measure of interest on validation data. In this paper we focus on the F-measure, which is widely used in multi-label problems. We derive two properties of the micro-averaged F measure, viewed as a function of the threshold values, which allow its global maximum to be found by an optimisation strategy with an upper bound on computational complexity of O(n^2 N^2), where N and n are respectively the number of classes and of validation samples. So far, only a suboptimal threshold selection rule and a greedy algorithm without any optimality guarantee were known for this task. We then devise a possible optimisation algorithm based on our strategy, and evaluate it on three benchmark, multi-label data sets.

Keywords: Pattern Recognition for Bioinformatics, Classification and Clustering, Machine Learning and Data Mining
Abstract: We present an approach for customized heartbeat classification of electrocardiogram (ECG) signals, based on the construction of one general multi-class classifier and one specific two-class classifier. The general classifier is trained on a global training dataset, containing examples of all possible classes and patterns. On the other hand, the individual-specific classifier is built using a small amount of individual data, which is a binary one-against-the-rest classifier, providing discrimination between normal and abnormal patterns from that individual. Such an individual-specific classifier can be a two-class classifier or a one-class classifier, depending on the availability of abnormal patterns in the individual training dataset. The classifications from the two classifiers are fused to obtain a final decision. The proposed approach is applied to the study of ECG heartbeat classification problem, significantly outperforming state-of-the-art methods. The proposed method can also be useful in anomaly detection of other biomedical signals.

Keywords: Biometrics
Abstract: Gender is one of the most important demographic attributes of human beings, and recently automatic face based gender classification has received increasing attentions due to its wide potential in many useful applications. To address such an issue, in this paper, we propose a novel variant of Local Binary Patterns (LBP), namely Local Circular Patterns (LCP). LCP makes use of clustering-based quantization instead of the binary coding strategy of the LBP operator, leading to an improvement in discriminative power. Meanwhile, thanks to the nature property of clustering-based quantization, LCP is more robust than LBP to noise. Experiments are carried out on the FERET database and the classification accuracy is up to 95.36%, clearly highlighting the effectiveness of the proposed method.

Keywords: Scene Understanding
Abstract: Studies on human faculties of scene recognition have lead to two broad classifications of the perceived information: local and global. It has been shown that both are processed separately and combined towards final category assignment. Recently, it was suggested that accuracy of computational models for local information closely match human performance, while it is not so for current global representations. In this paper, we propose a new global representation, AGIR. The key differences we propose to current approaches is the explicit modeling of `arrangement' (co-occurrence and configuration) in the scene, and use of multiple hierarchical dictionaries. The effectiveness of the proposed scheme is shown through various experiments and comparisons on both indoor and outdoor scene recognition.

Keywords: Features and Image Descriptors, Pattern Recognition for Bioinformatics, Image and Video Processing
Abstract: This paper presents the different steps for an automatic fluorescence-labelled cell classification method. First a data features study is discussed in order to describe cell texture by means of morphological and statistical texture descriptors. Then, results on supervised classification using logistic regression, random forest and neural networks, for both morphological and statistical descriptors, is presented. We propose a final consolidated classifier based on a weighted probability for each class, where the weights are given by the empirical classification performances. The method is evaluated on ICPR�12 HEp-2 dataset contest.

Keywords: Motion, Tracking and Video Analysis, Gesture and Behavior Analysis
Abstract: We propose a parallelized Annealed Particle Filter method via heterogeneous computing (P-APF), to implement real-time marker-less motion tracking based on OpenCL framework. The overall computing procedure in P-APF is decomposed into several computational tasks with corresponding granularity. According to the degree of parallelism, the tasks are assigned to standard and attached processors respectively, to fully leverage heterogeneous computing ability. A novel task latency hidden strategy is proposed to further reduce time cost. Experiments on different human motion datasets demonstrate that P-APF can achieve real-time tracking performance without losing accuracy. With an average acceleration ratio of 106 compared to serial implementation, the time cost basically remains constant with the growth of particle number and view number in a limited range.

Keywords: Pattern Recognition for Surveillance and Security
Abstract: In this paper we present a technique for real-time face logging in video streams. Our system is capable of detecting faces across a range of poses and of track- ing multiple targets in real time, grabbing face images and evaluating their quality in order to store only the best for each detected target. An advantage of our ap- proach is that we qualify every logged face in terms of a quality measure based both on face pose and on resolu- tion. Extensive qualitative and quantitative evaluation of the performance of our system is provided on many hours of realistic surveillance footage captured in dif- ferent environments. Results show that our system can simultaneously minimizing false positives and identity mismatches, while balancing this against the need to obtain face images of all people in a scene.

Keywords: Features and Image Descriptors, Neural Networks
Abstract: Recently the improved bag of features (BoF) model with locality-constrained linear coding (LLC) and spatial pyramid matching (SPM) achieved state-of-the-art performance in image classification. However, only adopting SPM to exploit spatial information is not enough for satisfactory performance. In this paper, we use hierarchical temporal memory (HTM) cortical learning algorithms to extend this LLC & SPM based model. HTM regions consist of HTM cells are constructed to spatial pool the LLC codes. Each cell receives a subset of LLC codes, and adjacent subsets are overlapped so that more spatial information can be captured. Additionally, HTM cortical learning algorithms have two processes: learning phase which make the HTM cell only receive most frequent LLC codes, and inhibition phase which ensure that the output of HTM regions is sparse. The experimental results on Caltech 101 and UIUC-Sport dataset show the improvement on the original LLC & SPM based model.

Keywords: Gesture and Behavior Analysis, Human Computer Interaction
Abstract: One of the most critical limitations of Kinecttexttrademark-based interfaces is the need for persistence in order to interact with virtual objects. Indeed, a user must keep her arm still for a not-so-short span of time while pointing at an object with which she wants to interact. The most natural way to overcome this limitation and improve interface reactivity is to employ a vision module able to recognize simple hand poses (e.g. open/closed) in order to add a state to the virtual pointer represented by the user hand. In this paper we propose a method to robustly predict the status of the user hand in real-time. We jointly exploit depth and RGB imagery to produce a robust feature for hand representation. Finally, we use temporal filtering to reduce spurious prediction errors. We have also prepared a dataset of more than 30K depth-RGB image pairs of hands that is being made publicly available. The proposed method achieves more than 98% accuracy and is highly responsive.

Keywords: Features and Image Descriptors, Classification and Clustering, Multimedia Analysis, Indexing and Retrieval
Abstract: In this paper, we propose a compact image signature based on VLAT. Our method integrates spatial information while significantly reducing the size of original VLAT by using two pojection steps. we carry out experiments showing our approach is competitive with state of the art signatures.

Keywords: Features and Image Descriptors, Motion, Tracking and Video Analysis
Abstract: This paper presents the development of the Tree BAsis Sparse-coding Inspired Similarity feature descriptor (TreeBASIS). TreeBASIS utilizes a binary vocabulary tree that is computed off-line using basis dictionary images (BDIs) derived from sparse coding, and the resulting tree is stored in memory for on-line searching. During the on-line algorithm, a small region around a feature point is passed into the BASIS tree, where a Hamming distance is computed between the region and the effectively descriptive BDI (EDBDI) to determine the branch taken. The path the FRI takes is saved as the descriptor, and matching is performed by following the paths of two features. Experimental results show that the TreeBASIS descriptor outperforms SIFT and SURF on frame-to-frame aerial feature point matching.

Keywords: 2D/3D Object Detection and Recognition, Vision for Robotics, Feature Reduction and Manifold Learning
Abstract: The advent of inexpensive depth augmented color (RGBD) sensors has brought about a large advancement in the perceptual capability of vision systems and mobile robots. Challenging vision problems like object category, instance and pose recognition have all benefited from this recent technological advancement. In this paper we address the challenging problem of pose recognition using simultaneous color and depth information. For this purpose, we extend a state-of-the-art regression framework by using a multi-kernel approach to incorporate depth information to perform more effective pose recognition on table-top objects. We do extensive experiments on a large publicly available dataset to validate our approach. We show significant performance improvements (more than 20%) over published results.

Keywords: Motion, Tracking and Video Analysis, Scene Understanding, Segmentation, Color and Texture
Abstract: Figure-ground labeling is a classical problem in computer vision in which the goal is to label different parts of the visual input as figural or background. Yet most existing approaches focuses on single image figure-ground labeling with little emphasis on video. We present a method which integrates several cues to achieve figure-ground labeling on video sequences. The method is evaluated on challenging video sequences.

Keywords: Features and Image Descriptors
Abstract: Abstract A novel extension for color images of the local phase quantization (LPQ) local descriptor is presented. The descriptor is obtained by using a multivector representation of color values in order to derive blur-robust features in frequency domain. We tested the proposed descriptor in texture classification problems, and quantified its robustness for several amounts of blur. The experiments show that the proposed descriptor achieves superior accuracy over its grayscale counterpart and other color texture descriptors. Furthermore its illumination-invariance properties guarantee remarkable performances in challenging scenarios of varying illumination, without the need of pre-processing textures with color-constancy algorithms.

Keywords: Stereo and Image-Based Modeling, Vision for Robotics, Vision for Graphics
Abstract: This paper presents a new stereo-motion approach for 3D scene reconstruction in dense and accurate form, that allows the cameras to be described by the full perspective model. Given a short and arbitrary motion of a stereo rig of camera, the projective depth of every image point can be recovered from the rank-four property of a matrix that comprises the image positions of the scene, and the associated 3D position can thereby be determined accurately. Compared to earlier methods, the approach allows the use of the full perspective model to describe the cameras, and thus can attain a higher accuracy. In addition, the projective depths are recovered without the need of having initial guess of the depth map or going through iterations or approximation. The recovery process demands only a few stereo correspondences over the entire scene to start with, and points that are occluded in some of the views can also be reconstructed. Experiments on real image sequences are shown to illustrate the effectiveness of the approach.

Keywords: 2D/3D Object Detection and Recognition, Machine Learning and Data Mining
Abstract: Automatic 3D reconstruction of human poses from monocular images is a challenging and popular topic in the computer vision community, which provides a wide range of applications in multiple areas. Solutions for 3D pose estimation involve various learning approaches, such as support vector machines and Gaussian processes, but many encounter difficulties in cluttered scenarios and require additional input data, such as silhouettes, or controlled camera settings. We present a framework that is capable of estimating the 3D pose of a person from single images or monocular image sequences without requiring background information and which is robust to camera variations. The framework models the non-linearity present in human pose estimation as it benefits from flexible learning approaches, including a highly customizable 2D detector. Results on the HumanEva benchmark show how they perform and influence the quality of the 3D pose estimates.

Keywords: Motion, Tracking and Video Analysis
Abstract: Accurate segmentation provides a useful contour constraint to alleviate drifting during online learning for tracking. Towards this end, we present a closed-loop approach for object tracking that links Hough forests and alpha matting via an effective back-projection scheme for patches. A novel hybrid-Hough-forests-based method first estimates object location. Given the object location, the trimap of matting is then automatically generated from the patches back-projected from the Hough forests. Subsequently, an accurate contour of the object can be obtained based on a robust matting technique. Based on such an accurate contour, an update strategy is utilized to obtain reliable labeled samples to update the Hough forests to decrease the risk of model drift. Extensive comparisons on challenging sequences demonstrate the robustness and effectiveness of the proposed method.

Keywords: 2D/3D Object Detection and Recognition
Abstract: This paper describes an approach to location and orientation estimation of a person's face with color image and depth data from a Kinect sensor. The combined 2D and 3D histogram of oriented gradients (HOG) features, called RGBD-HOG features, are extracted and used throughout our approach. We present a coarse-to-fine localization paradigm to obtain localization results efficiently using multiple HOG filters trained in support vector machines (SVMs). A feed-forward multi-layer perception (MLP) network is trained for fine face orientation estimation over a continuous range. The experimental result demonstrates the effectiveness of the RGBD-HOG feature and our face pose estimation approach.

Keywords: Features and Image Descriptors, Performance Evaluation
Abstract: Pedestrian detection problem has been a touchstone of various image feature descriptors. In this paper, we evaluate four kinds of representative local descriptors (HOG, Haar-like, SURF and LBP) for pedestrian representation. Our goal is to find out the best combination of feature descriptors by analyzing and evaluating the complementarities of them. With the cross validation method, we first find out the best descriptor, which is then combined with other descriptors one by one for evaluation. In addition to direct descriptor combination, we propose a new descriptor strategy, called structural combination. Experiments on two public pedestrian datasets show that the performance evaluation can support the complementarily analysis and the complementarities is relevant to combination strategies.

Keywords: Features and Image Descriptors, Segmentation, Color and Texture, Multimedia Analysis, Indexing and Retrieval
Abstract: Texture analysis algorithms are employed in many computer vision applications. A group of high performing texture algorithms are based on the concept of local binary patterns (LBP) which describe the relationship of pixels to their local neighbourhood. LBP descriptors are invariant to intensity changes and rotation invariance is simple to derive. In addition, LBP features can be calculated for different neighbourhood radii and thus allow texture description at different scales. In conventional LBP methods, the histograms corresponding to different radii are simply concatenated which results in a loss of information between these scales and added ambiguity. In this paper, we address this problem and show that multi-dimensional LBP histograms provide effective texture descriptors. We demonstrate, on various texture datasets from the Outex suite and both for texture classification and texture retrieval scenarios, that our proposed approach consistently outperforms conventional LBP features.

Keywords: Low-Level Vision, Motion, Tracking and Video Analysis, Geometric and Photometric Registration
Abstract: The underlying principle behind most optical flow algorithms is that the brightness of a pixel remains the same as it flows from one frame to the next. The first order Taylor approximation used in formulating this brightness constancy principle may not be accurate when intensity profiles change non-linearly. In this paper, we propose a method of alleviating the effect of this approximation. Instead of computing image gradients using conventional horizontal and vertical filters of fixed coefficients and sizes, we propose to obtain the gradient information by an explicit examination of ramp profiles at a given location, in all directions. The gradient information obtained using the proposed analysis is more robust under non-linear changes in intensity profiles. Our results demonstrate that by incorporating the ramp structure information as proposed, we are able to improve existing optical flow algorithms.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Surveillance and Security
Abstract: This paper presents a novel method for pedestrian counting in surveillance videos, which localizes and tracks the head-shoulders of pedestrians via the integrated bottom-up/top-down processes. In the bottom-up stage, we extract and match informative local image features crossing frames to obtain the initial moving regions (i.e. potential pedestrians). The top-down stage comprises two steps: (i) head-shoulder verification via a part-based classifier and (ii) head-shoulder tracking guided by the motion and appearance consistency. Moreover, the geometric context of the camera is employed to effective narrow the searching space of inference. We apply the method with the challenging videos and outperform the state-of-the-arts approach.

Keywords: 2D/3D Object Detection and Recognition, Remote Sensing, Pattern Recognition for Surveillance and Security
Abstract: Detection of moving vehicles in wide area motion imagery (WAMI) is increasingly important, with promising applications in surveillance, traffic scene understanding and public service applications such as emergency evacuation and policy security. However, the large camera motion, along with low contrast between vehicles and backgrounds, makes detection a challenging task. In this paper, we propose a novel moving vehicle detection approach by embedding the scene context, which is a road network estimated online. A two-step framework is used in the work. First, with an initial vehicle detection, trajectories are achieved by vehicle tracking. Then, the road network is extracted and used to reduce false detections. Quantitative evaluation demonstrates that the proposed contextual model remarkably improves the detection performance.

Keywords: Segmentation, Color and Texture, 2D/3D Object Detection and Recognition, Detection, Separation and Segmentation
Abstract: In this work, we propose a novel segmentation method based on the continuous max-flow (CMF) formulation of the Potts model incorporating the statistical shape model. We increase the robustness and accuracy of the Potts model by using the prior shape knowledge from the Principal Component Analysis (PCA) to represent the desired shape. Our multi-label model can segment several objects simultaneously and guarantee one label with the structure similar to the shape prior. The proposed approach is applied to both synthetic and medical image of liver from computed tomography (CT). Numerous numerical experiments demonstrate that our model is efficient and with good quality in practice.

Keywords: Motion, Tracking and Video Analysis, Medical Image Analysis and Registration, Segmentation, Color and Texture
Abstract: The study of neurological processes and pharmaceutical effects often relies on the analysis of mice behaviour. Automatic tracking tools are widely employed for this purpose, however they are mainly limited to a single mouse. We propose a real time segmentation and tracking algorithm able to manage multiple interacting mice regardless of their fur colour or light settings via an infrared camera. The approach proposed combines position, temperature and shape information thanks to the two main contributions of this paper: the ``temporal watershed'' and its information fusion with mice ``heat signatures''. The former segments shapes thanks to an extension of a classical seed-based segmentation algorithm in a expectation maximization framework; the latter contributes in mice identities preservation through the dynamic heat distribution of each body. Preliminary results show that our algorithm achieves performance comparable to the state of art, even with a larger number of targets to be tracked.

Keywords: 2D/3D Object Detection and Recognition, Features and Image Descriptors, Statistical, Syntactic and Structural Pattern Recognition
Abstract: In this paper, we propose a novel approach for object detection via foreground feature selection and part-based shape model. It automatically learns a shape model from cluttered training images without need to explicitly given bounding box on objects. Our approach commences by extracting a set of feature descriptors, and iteratively selects the foreground features using Earth Movers Distances based matching. This lead to a part-based shape model that can be used for object detection. Experimental results show that the proposed method has comparable performance with the state-of-the-art shape-based detection methods but with less requirements on the data at the training stage.

Keywords: Motion, Tracking and Video Analysis, Image and Video Understanding
Abstract: Millions of video surveillance cameras distribute around the world, and capture tremendous number of video data endlessly. Video browsing by frame are time consuming and inefficient, since needless information is abundant in the raw videos. Video synopsis is an effective way to solve this problem by producing a short video representation, while keeping the essential activities of the original video. However, traditional video synopsis only eliminates redundancy in spatial and temporal domain, while neglects redundancy in content domain. However, too many observations will make synopsis video confusing and degrade the subjective effect of synopsis video. In this paper, we present a novel video synopsis method based on key observation selection. Key observation selection is conducted for activity to eliminate content redundancy. We have demonstrated the effectiveness of our approach on real surveillance videos.

Keywords: Motion, Tracking and Video Analysis
Abstract: In this paper, we propose a new method to track players using 3D particle filter guided by the time-situation graph in order to perform players tracking robust to occlusion in a soccer image sequence. In the conventional method using particle filter, there is a deficit that it is difficult to discover the players again once they are lost in an image sequence. Thus, we represents the position information of two or more players as the time-situation graph beforehand. Then, by running particle filter guided by this graph, the incorrect detection of players can be greatly reduced and the players be robustly tracked even when occlusion occurs. As a result, the tracking accuracy was improved by 7.15 points in comparison with the conventional method.

Keywords: Features and Image Descriptors, Vision for Robotics
Abstract: The paper introduces a novel approach to place representation for robot localization and mapping. It uses classical invariance theory while proposing an adaptive kernel to omnidirectional images and exploiting only the main significant visual information in the images. The approach is validated in real world robot exploration and localization and compared to color histograms.

Keywords: Features and Image Descriptors, Classification and Clustering, Pattern Recognition for Surveillance and Security
Abstract: We consider the problem of extracting a signature representation of similar entities employing covariance descriptors. Covariance descriptors can efficiently represent objects and are robust to scale and pose changes. We posit that covariance descriptors corresponding to similar objects share a common geometrical structure which can be extracted through joint diagonalization. We term this diagonalizing matrix as the Covariance Profile (CP). CP can be used to measure the distance of a novel object to an object set through the diagonality measure. We demonstrate how CP can be employed on images as well as for videos, for applications such as face recognition and object-track clustering.

Keywords: Motion, Tracking and Video Analysis, Vision for Robotics
Abstract: Calibrating hand-eye geometry is often based on explicit feature correspondences. This article presents an alternative method that uses the apparent flow induced by the motion of the camera to achieve selfcalibration. To make the method more robust against noise, the strategy is to use the orientation of the normal flow field which is more noise-immune, to recover first the direction component of the hand-eye geometry. Outliers in the extracted flow data are identified using some intrinsic properties of the flow field together with the partially recovered hand-eye geometry. The final complete solution is refined using a robust process. The proposed method can also be used for determining the relative geometry of multiple cameras without demanding overlap in the visual fields of the cameras. Experimental results on synthetic data and real image data are shown to illustrate the feasibility of the method.

Keywords: Features and Image Descriptors, 2D/3D Object Detection and Recognition
Abstract: In this paper we propose an original framework for the description and the subsequent recognition of objects of limited size. Although of general applicability, the framework is presented here as a way to trace different yet similar metal tools employed in the mechanical constructions industry. For the purpose of object description, time-varying silhouettes of the object are acquired under turntable motion and collated into a single image. The resulting footprint matrix represents the object in both a compact and effective way. Visual matching of footprint matrices is carried out with a computationally efficient algorithm that is organized into three distinct levels so as to benefit of a progressive suppression of the irrelevant information.

Keywords: Low-Level Vision, Speech and Audio Analysis, Scene Understanding
Abstract: Using multiple receivers (microphones or antennas) in a rigid configuration, such as on a smartphone, it is possible to measure time difference of arrival to the receivers. This in turn can be used to determine the direction to the transmissions, if there are at least three receivers. When using two receivers it can be used to determine the angle to the transmissions relative to the line through the two receivers. In this paper we study three minimal problems for pose using such data: (i) determine position and orientation using five transmissions, (ii) determine position and orientation using four transmissions and known 'down' direction and (iii) determine position using three transmissions and known orientation. Numerically stable solvers are implemented. An experimental validation of the solvers are performed on simulated data.

Keywords: Segmentation, Color and Texture
Abstract: Segmenting objects from color images to obtain useful information is a challenging research area recently. In this paper, a novel algorithm by combining a saliency map with an extension of a geometric active contour model is proposed to automatically segment the object of interest. The saliency map is first generated from the input image by a histogram based contrast method. The most salient regions are then detected as dominant parts of the object. After that, a contour is initialized using salient regions determined. Finally, by applying a geometric active contour model, the contour starts evolving iteratively to segment object boundaries. Experimental results attained on various natural scene images have shown that our proposed method is able to not only replace manual initialized contour and improve the accuracy, noise robustness of segmentation but converge to an optimal solution earlier than recent active contour models as well.

Keywords: Motion, Tracking and Video Analysis, Image and Video Understanding, Image and Video Processing
Abstract: This paper proposes a novel accident prediction approach based on extracting the relation between interested vehicles and increasing risk factor according to anomaly detection in real time traffic videos. In learning process of the traffic model at intersections, we detect all trajectories by tracking of each vehicle and then group them considering road model. All trajectories are clustered by Continuous Hidden Markov Model with Mixture of Gaussian (MoG) and Common Route Model (CRM) for each group of trajectories is found. After extracting all CRM�s and defining their relations, in real time traffic analysis process, partial motion of the vehicles are evaluated and anomalies are detected if there is. In this approach, while searching for accident risk, partial trajectories of vehicles are classified to the most similar CRM�s. For each source vehicle, risk factors are calculated with target vehicles that are in related CRM�s and has Region of Interest (ROI) intersected with source vehicle. The advantage of this approach is that the system does only analyze vehicles in accident risk and this increases the performance of the system. Beside these, since CRM information and their features like relations, directions and likelihood in classification process are learned, anomalies can easily be detected and used as risk enhancer. Experimental results show that the proposed model has high prediction rate in real world accident events.

Keywords: Stereo and Image-Based Modeling, Motion, Tracking and Video Analysis
Abstract: Representing a video by a set of key frames is useful for efficient video browsing and retrieving. But key frame extraction keeps a challenge in the computer vision field. In this paper, we propose a joint framework to integrate both shot boundary detection and key frame extraction, wherein three probabilistic components are taken into account, i.e. the prior of the key frames, the conditional probability of shot boundaries and the conditional probability of each video frame. Thus the key frame extraction can be treated as a Maximum A Posteriori problem which can be solved by adopting alternate strategy. Experimental results show that the proposed method preserves the scene level structure and extracts key frames that are representative and discriminative.

Keywords: Motion, Tracking and Video Analysis, Occlusion and Shadow Detection, Segmentation, Color and Texture
Abstract: In this paper, we address the problem of detecting occlusion boundaries from video sequences. We build a bi-directed graph whose nodes are line fragments extracted from superpixels's edges. Based on the graph, we compute a global occlusion saliency map by integrating motion, shape and topology cues into the framework of Saliency Network. Furthermore, with the structural information generated from the network, the property of structural consistency is proposed to prune the graph and refine the saliency map. Finally, we train a classifier to detect occlusion fragments combining the global saliency value and local edge strength. The detector outperforms the state-of-the-art on the benchmark of Stein and Hebertcite{Stein09occlusionboundaries} by improving average precision to .80.

Keywords: Features and Image Descriptors, 2D/3D Object Detection and Recognition, Image and Video Understanding
Abstract: Visual cortex inspired features mimic what we know of the brain's visual cortex, which is still the best existing object detection system regarding speed and accuracy. For this paper we benchmarked two prominent implementations of these features, Mutch and Lowe's SLF-HMAX and Pinto et al.'s V1-like, against the popular local invariant features SIFT and PHOW in combination with the bag of visual words approach. The benchmark task was the detection of illicit objects in X-ray images of luggage. X-ray inspection is one of the main means of preventing the transport of illicit objects into sensitive areas. The visual cortex inspired features performed superior to the conventional features, probably owing to the textureless nature of X-ray images and the encoding of geometric information.

Keywords: Features and Image Descriptors, Classification and Clustering, Biometrics
Abstract: In this paper, we propose a fully automatic approach for person-independent 3D facial expression recognition. In order to extract discriminative expression features, each aligned 3D facial surface is compactly represented as multiple global histograms of local normal patterns from multiple normal components and multiple binary encoding scales, namely Multi-Scale Local Normal Patterns (MS-LNPs). 3D facial expression recognition is finally carried out by modeling multiple kernel learning (MKL) to efficiently embed and combine these histogram based features. By using the SimpleMKL algorithm with the chi-square kernel, we achieved an average recognition rate of 80.14% based on a fair experimental setup. To the best of our knowledge, our method outperforms most of the state-of-the-art ones.