Keywords: Computer-Aided Diagnosis and Surgery, Enhancement, Restoration and Filtering
Abstract: To enhance ICP monitoring of Traumatic Brain Injury (TBI) patients, much research effort has been attracted to the development auto-alarming systems and forecasting methods to predict impending intracranial hypertension episodes. Nevertheless, the performance of the proposed methods are often limited by the presence of artifacts in the ICP signal. To address this bottleneck, we propose novel artifact correction methods. A scale-based filter is pro- posed to identify the artifacts. For the proposed filter, instead of classic statistics, robust statistics is employed to estimate the scale parameter. Thus, our proposed methods are robust against undesirable influences from artifacts. Since the ICP signal is non-stationary, non-stationary signal processing techniques, the empirical mode decomposition (EMD), wavelet transformation and median filter, are also employed. The effectiveness of proposed methods are evaluated experimentally. Experimental results demonstrate that, with the proposed artifact correction methods, significant performance gains can be achieved.

Keywords: Medical Image Analysis and Registration, Image and Video Processing
Abstract: Accurate vessel segmentation is the first step in retinal image analysis for medical diagnosis. In this paper we propose a novel method to segment vessel network in fundus image. Vessel centerlines are first extracted by using a set of directional line detectors. Next an Iterative Geodesic Time Transform (ItGTT) is designed to segment the entire vessel network. The idea of the ItGTT is to use centerline pixels as the initial reference set, and compute geodesic time between candidate vessel pixels and the reference set iteratively with an adaptive reference set updating strategy. The entire vessel network is the binary image of the last reference set at the end of the iteration. Experimental results demonstrate that our method can segment retinal vessels effectively.

Keywords: Human Computer Interaction, Gesture and Behavior Analysis, 2D/3D Object Detection and Recognition
Abstract: Hand posture recognition (HPR) plays an important role in human-computer interaction (HCI) since it is one of the most common and natural ways of communication among human beings. Different fingers often represent different meanings which will attract more attentions in HPR research. Based on finger geometric feature and its classification, we develop a HPR system that can tell its posture on possible fingers. We explore kinematic constraints of the hand with forearm to extract finger geometric features which are translation, rotation and scale invariant. We first search hand components with the help of skeleton, and then order them into a serial arrangement according to either left hand or right hand and extract the geometric features among fingers, palm and forearm, finally those features are used in SVM classification for HPR. Our method can recognize twelve different types of hand postures for both hands respectively. Experiments under different illumination conditions and different scenes demonstrate the effectiveness and efficiency of the proposed method.

Keywords: Classification and Clustering, Machine Learning and Data Mining, Feature Reduction and Manifold Learning
Abstract: Hierarchical classification, decomposing the multi-class classification problem into binary ones hierarchically, is efficient when the class quantity getting large. Nowadays, the variety of features to describe data becomes huge and meanwhile the form of these features is diverse, which both make the task of feature fusion crucial for classification. In this paper, an adaptive kernel learning method, which resorts to kernel combination for feature fusion, is proposed and incorporated into the hierarchical classification framework for the multi-class and multi-feature classification scenario. By the centered kernel alignment, the tasks of category partition and kernel combination are unified into a coherent optimization problem, and an iterative algorithm is de-signed to solve it. Experimental results on two datasets show that our method is not only efficient but also accurate compared with other baseline methods.

Keywords: Classification and Clustering
Abstract: We propose a novel method for automatic detection of the transport mode of a person carrying a Smartphone. Existing approaches assume idealized positioning data with no GPS signal losses, require information from additional external sources such as real time bus locations, or only allow for a coarse distinction between very few categories (e.g. �still�, �walk�, �motorized�). Our approach is designed to deal with cluttered real-world Smartphone data and can distinguish between fine-grained transport mode categories. It is robust against GPS signal losses by including positioning data obtained from the cellular network and data from accelerometer readings. Mode detection is performed by a two-stage classification technique using randomized ensemble of classifiers combined with a Hidden Markov Model. We report promising results of an experimental performance analysis with real-world data collected by 15 volunteers during their everyday routines over a period of two months.

Keywords: Machine Learning and Data Mining, Handwriting Recognition
Abstract: In likelihood ratio based decision methods, often a variable number of input evidences is used. A decision based on many such inputs can result in nearly the same likelihood ratio as one based on few inputs. We consider methods for distinguishing between such situations. One of these is to provide confidence intervals together with the decisions and another is to combine the inputs using weights. We suggest a new method that generalizes Bayesian approach and uses an explicitly defined discount function. Empirical evaluation shows greater flexibility of the designed method.

Keywords: Machine Learning and Data Mining, Neural Networks, Classification and Clustering
Abstract: Spatial pyramid matching (SPM) component is part of most state-of-art image classification methods. SPM encodes spatial distribution of image features, in an unsupervised fashion, by partitioning an image into regions at multiple scales and concatenating feature vectors for these regions. In this paper we propose to replace the unsupervised SPM procedure with a supervised two-stage feature selection that requires the image partitioned at a single scale. Experimental results show the proposed method performs statistically significantly better than the SPM baseline.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition, Classification and Clustering
Abstract: Learning from large, multi-class data sets poses great challenges to ensemble methods. The weak learner condition makes the conventional method inappropriate to handle multi-class classification, which leads to early termination of the training process. Also, elongated training time makes learning from large data set infeasible. To circumvent these issues, we present a novel method that integrates a sampling strategy and an error parameter that alters weighted error. Experiments were conducted with ten data sets from real world applications. It is evident that our proposed method achieves greater performance and avoids early termination. In addition, our method significantly improves training efficiency and accommodates training with large data set.

Keywords: Classification and Clustering, Neural Networks, Feature Reduction and Manifold Learning
Abstract: Motivated by the fact that data of each cluster are often well captured by distinct features, we propose a clustering approach called multiple kernel self-organizing map (MK-SOM) that integrates multiple kernel learning into the learning procedure of SOM, and carries out cluster-dependent feature selection simultaneously. MK-SOM is developed to reveal the intrinsic relation between features and clusters, and is derived with an efficient optimization procedure. The proposed approach is evaluated on two benchmark datasets, UCI and Caltech-101. The promising experimental results demonstrate its effectiveness.

Keywords: Pattern Recognition for Bioinformatics, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: Efficiency of facial feature detection is very crucial in face related applications such as face recognition and reconstruction. Traditional algorithms of high precision are often with expensive computation. In this paper, we propose a fast facial feature detection algorithm with good precision. The basic idea is to combine fast search strategy in the global image and high precision classifier in the local regions. In a global image, we borrow the idea of Active Shape Model (ASM) and utilize the average search template to decrease the search area in images. At each local region, we use trained random forest classifier (RF) to identify the existence of facial feature. An iteration procedure of template adjustment is specially designed to ensure the detection precision. Experiments show the effectiveness of the proposed facial feature detection algorithm.

Keywords: Gesture and Behavior Analysis, Pattern Recognition for Surveillance and Security, Motion, Tracking and Video Analysis
Abstract: Human action recognition is an active area with applications in several domains such as visual surveillance, video retrieval and human-computer interaction. Current approaches assign action labels to video streams considering the whole video as a single image sequence. Such approaches, albeit very refined, may fail on some samples due to large variability between frames, suggesting that features extracted from training videos may adopt description and classification models that may better represent video portions rather than the whole stream. To this aim, in this paper we propose a multiple subsequence combination method dividing the video into a number of consecutive subsequences and classifying each one applying a part-based method in conjunction with the bag of visual words approach. We classify a video combining subsequence labels according to rules inspired by the multiple expert system framework. We extensively tested our approach on the KTH, UCF sport and Youtube datasets showing, on the one side, that it outperforms a method classifying action using the whole stream and, on the other side, that its performance are comparable with the best published ones for each dataset.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, Multimedia Analysis, Indexing and Retrieval, 2D/3D Object Detection and Recognition
Abstract: In this paper, we propose a 3D model retrieval system using sketch queries. The 3D model is described by a single characteristic view with the biggest exposure opportunity. The sketch query and the characteristic view are divided into several parts. Each part is quantized into a Fourier descriptor, and the spatial arrangements of these parts are measured by the graph spectra. Our contributions are as follows: (1) We propose a new characteristic view, namely, the Biggest Exposure Opportunity View(BEOV) to represent a 3D model; (2) The details inside the sketch query and the BEOV are considered. Experiments were evaluated using the Princeton shape benchmark data sets. Our approach improves the retrieval precision and reduces the computational cost when compared with the National Taiwan University~(NTU) 3D model search engine.

Keywords: Gesture and Behavior Analysis
Abstract: Automatically identifying and analyzing head gestures is useful in many situations like smart meeting rooms and intelligent driver assistance. In this paper, we show that head movements can be broken into its elemental forms (i.e. moving and fixation states) and combinations of these elemental forms give rise to various head gestures. Our approach which we term, Optical flow based Head Movement and Gesture Analyzer (OHMeGA), segments head gestures into moving and fixation states using optical flow tracking and intermittent head pose estimation. OHMeGA runs in real-time, is simple to implement and set up, is robust and is accurate. Furthermore, segmenting head gestures into its elemental forms gives access to higher level semantic information such as fixation time and rate of change of head motion. Experimental analysis shows promising results.

Keywords: Classification and Clustering, Pattern Recognition for Surveillance and Security
Abstract: We propose an approach to recognize group activities which involve several persons based on modeling the interactions between human bodies. Benefitted from the recent progress in pose estimation , we model the activities as the interactions between the parts belong to the same person (intra-person) and those between the parts of different persons (inter-person). Then a unified, discriminative model which integrates both types of interaction is developed. The experiments on the UT-Interaction Dataset show the promising results and demonstrate the power of the interaction models.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, Detection, Separation and Segmentation, Performance Evaluation
Abstract: Music listeners often mishear the lyrics to unfamiliar songs heard from public sources. Speech recognition and acoustic distance can be used to solve this problem. This paper proposes a robust lyric search method for music information retrieval. Firstly, weighted syllable confusion matrix (WSCM) is derived from confusion network. And then, we apply acoustic distance, which is computed based on a WSCM of an automatic speech recognition (ASR) experiment, into dynamic programming (DP) based syllable string matching. The experimental results show that this method increases the hit rate compared with the conventional acoustic distance and edit distance.

Keywords: Character and Text Recognition, Detection, Separation and Segmentation, Layout Analysis
Abstract: An automatic text recognizer needs, in first place, to localize the text in the image the more accurately possible. For this purpose, we present in this paper a robust method for text detection. It is composed of three main stages: a segmentation stage to find character candidates, a connected component analysis based on fast-to-compute but robust features to accept characters and discard non-text objects, and finally a text line classifier based on gradient features and support vector machines. Experimental results obtained with several challenging datasets show the good performance of the proposed method, which has been demonstrated to be more robust than using multi-scale computation or sliding windows.

Keywords: Character and Text Recognition, Classification and Clustering, Image and Video Understanding
Abstract: Reading text from scene images is a challenging problem that is receiving much attention, especially since the appearance of imaging devices in low-cost consumer products like mobile phones. This paper presents an easy and fast method to recognize individual characters in images of natural scenes that is applied after an algorithm that robustly locates text on such images. The recognition is based on a gradient direction feature. Our approach also computes the output probability for each class of the character to be recognized. The proposed feature is compared to other features typically used in character recognition. Experimental results with a challenging dataset show the good performance of the proposed method.

Keywords: Document Understanding, Pattern Recognition for Surveillance and Security
Abstract: We consider a new text classification task: classifying enterprise email messages into sensitive business topics. The identification of sensitive topics in email messages is important for enterprises to safe-guard their critical data such as intellectual properties and trade secrets. We introduce the incremental PCA (Principal Component Analysis) to email representation, which can learn a feature subspace incrementally and effectively reduce the feature dimensionality. Linear SVM (Support Vector Machine) is then adopted to learn the classification models. We validate our approaches with 5,000 emails extracted from the Enron Email set. Experimental results show that SVM outperforms other classification methods, and the incremental PCA produces a substantial reduction in the processing time and a slight increase in the classification accuracy compared to SVM with all the features.

Keywords: Character and Text Recognition
Abstract: This paper presents a new discriminative linear regression approach to adaptation of a discriminatively trained prototype-based classifier for Chinese OCR. A so-called sample separation margin based minimum classification error criterion is used in both classifier training and adaptation, while an Rprop algorithm is used for optimizing the objective function. Formulations for both model-space and feature-space adaptation are presented. The effectiveness of the proposed approach is confirmed by experiments for adaptation of font styles and low-quality text, respectively.

Keywords: Character and Text Recognition, Camera-Based Document Analysis, Detection, Separation and Segmentation
Abstract: We propose a new component-tree based method with efficient and effective pruning strategies for user- intention guided text extraction from scene images. A grayscale image is represented first as two component- trees, whose nodes represent possible candidates of character components. The non-text candidates are then pruned by using contrast, geometric and text line information as well as the constraint imposed by user intention. Bounding boxes of the intended words are then generated by grouping the survived text objects. Finally, the better result is chosen as extracted text from the two trees using an evaluation function. The proposed algorithm is evaluated on ICDAR-2003 benchmark dataset and promising results are achieved.

Keywords: Document Analysis Systems
Abstract: A document image matching approach making use of probabilistic graphical models is proposed. The document image is first represented by a tree with the nodes in the tree corresponding to the regions in the image and the edges indicating the parent-child relationships between them, transforming the problem to tree matching. A graphical model, i.e. pairwise Markov Random Field is defined on the tree, in which sense the nodes are considered as random variables and the edges encode the relations among these variables in the probability domain. The tree matching problem is then formulated as Maximum a Posterior (MAP) inference over the graphical model and solved by belief propagation. Since the underlying graphical model is tree-structured, the exact inference can be obtained. With properly defined potential functions in the joint probability represented by the graphical model, the disparity in tree representations caused by different image capturing conditions can be tolerated as demonstrated in the encouraging experimental results.

Keywords: Performance Evaluation, Document Analysis Systems, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Repeated appearance of any block of spatial data in document images can be cached and encoded single time to get good compression ratio. This Reusable Document Component (RDC) can replicate the blocks of each redundant image at the receiver side at different positions but with same size and orientation. We have proposed a novel algorithm of Orientation Scale mapped RDC (OS-RDC) which can identify and cache repeated image blocks even if they are of different size and orientations. Both the inter-page and intra-page redundancies are addressed ensuring significant quality preservation.

Keywords: Document Understanding, Camera-Based Document Analysis, Graphics Recognition
Abstract: Teaching boards are omnipresent in classrooms throughout the world. Tableau is a software environment for processing images from teaching-boards acquired using portable digital cameras and cell-phones, being the first software environment able to process non-white boards. The aim of the Tableau environment goes far beyond processing and compressing teaching-board images, it also targets at content analysis and recognition. This paper presents the content analysis tool in Tableau that is able to automatically segment and classify textual and pictorial areas.

Keywords: Document Analysis Systems, Performance Evaluation, Character and Text Recognition
Abstract: This paper presents the current status of Decapod's English font reconstruction. The Potrace algorithm and its parameters that affect glyph shape are examined. The visual fidelity of Decapod's font reconstruction is shown and compared to Adobe ClearScantexttrademark. The font reconstruction details of the two methods are presented. The experiment demonstrates the capabilities of the two methods in reconstructing the font for a synthetic book typeset each time with one of six English fonts, three serif and three sans-serif. For both typefaces, Decapod is able to create a reusable TTF font that is embedded in the generated PDF document.

Keywords: Document Analysis Systems, Document Understanding
Abstract: Following the recent trend in using low level image features in classifying document images, in this paper we present a novel approach for structured document classification by matching the salient feature points between the query image and the reference images. Our method is robust to diverse training data size, image formats and qualities. Through matching the feature points, image registration is available for the query image as well. Although we aimed for the large domain of the structured document images, our method already achieved zero error rates in the tests on the benchmark NIST tax form databases.

Keywords: Handwriting Recognition, Character and Text Recognition
Abstract: This paper presents a handwritten digit recognition method based on cascaded heterogeneous convolutional neural networks (CNNs). The reliability and complementation of heterogeneous CNNs are investigated in our method. Each CNN recognizes a proportion of input samples with high-confidence, and feeds the rejected samples into the next CNN. The samples rejected by the last CNN are recognized by a voting committee of all CNNs. Experiments on MNIST dataset show that our method achieves an error rate 0.23% using only 5 CNNs, on par with human vision system. Using heterogeneous networks can reduce the number of CNNs needed to reach certain performance compared with networks built from the same type. Further improvements include fine-tuning the rejection threshold of each CNN and adding CNNs of more types.

Keywords: Handwriting Recognition, Document Understanding, Document Analysis Systems
Abstract: Handwritten identification is a technique of automatic person identification play important role in economic dispute case. In this paper, we propose a new dataset system of Economic Dispute handwritten (DSEDH) based on stroke shape and structure Features. The system consists of CIMS, CPIS, HMS, HIS Subsystem, segmentation of character, recognition of handwritten, this system with high efficiency and accuracy of compute. This system is running on integration of VC++, MATLAB. Database is Microsoft SQL Server. By this DSEDH system, management data of Economic Dispute handwritten can be obtained accurately. system offers a new efficient method for management of handwritten.

Keywords: Character and Text Recognition, Camera-Based Document Analysis, Image and Video Understanding
Abstract: Recent market report on Consumer Electronics shows that TV with Internet is going to be one of the most demanding products for the near future. As a consequence the demand of recognizing the TV context also increases among the research community. Lots of information about the TV video can be obtained from the meta data in case of digital TV broadcast. But in the developing countries like India, still the penetration of digital TV is only about 10%. So there is a need for video OCR system that can recognize the textual context of the TV video, since fair amount of contextual information can be obtained from the graphical text of the TV video. In this paper a low complexity video OCR system has been presented, that can be deployed on an embedded platform. The novelty of the proposed method is the use of low processing cycle and memory and yet getting a recognition accuracy of 84.23% which is higher than the usual video OCR recognition accuracy. Moreover, the proposed method can recognize about 180 characters on average per frame in 26.34 milliseconds.

Keywords: Graphics Recognition, Pattern Recognition for Search, Retrieval and Visualization
Abstract: Symbol retrieval into technical documents is still a hot challenge in the document analysis community. In this paper we propose another way to spot symbols. A pixel-based template operator which is an adaptation of the hit-or-miss transform is defined. This operator is robust to translation, rotation and reflection. Experimental results on a real application show the efficiency of our approach.

Keywords: Character and Text Recognition, Speech and Audio Processing, Image and Video Processing
Abstract: We propose a new Iterative-Midpoint-Method (IMM) for video character gap filling based on end pixels and neighbor pixels in the extracted contour of a character. The method obtains the Enhanced Gradient Image (EGI) for the given gray character image to sharpen text pixels. Max-Min clustering and K-means clustering algorithm with K=2 are applied on the EGI to obtain text candidates. To clean up the background information, the intersection of the text candidate image and the Sobel of the input image is considered. The method extracts edges in Canny of the input image corresponding to pixel in the intersection results, which we call the potential candidates having possible contour of the character with fewer disconnections. From the contour, we identify the correct pair of end pixels based on mutual nearest neighbor criteria. The three midpoints obtained from the two end pixels and their two preceded pixels are noted. The distance between three consecutive midpoints is used to predict a new midpoint. From the new midpoint, the method recursively computes midpoints till it reaches end pixels, which results in updated new end pixels. In this way, the method repeats midpoint computation iteratively to fill the complete gap between two end pixels. The method has been tested on 500 images which include 200 character images from video, 200 character images from ICDAR-2003 competition data and 100 images from object data to evaluate the performance. The comparative study shows that the proposed method is superior to a baseline method in terms of recognition rate.

Keywords: Character and Text Recognition
Abstract: This paper addresses the problem of recognizing rotated characters in scene and estimating the rotation angle using Weighted Direction Code Histogram (WDCH) and Modified Quadratic Discriminant Function (MQDF). In our previous paper, we proposed a rotation-free character recognition method and confirmed the feasibility of real world application. We also applied our method to scene analysis to detect surface normal from written characters by estimating rotation angle of the characters. In this research, we evaluate how precisely our proposed method can estimate rotation angle by comparing obtained result by 3D digitizer.

Keywords: Character and Text Recognition
Abstract: In this paper, we propose a novel text detection approach based on stroke width. Firstly, a unique contrast-enhanced Maximally Stable Extremal Region(MSER) algorithm is designed to extract character candidates. Secondly, simple geometric constrains are applied to remove non-text regions. Then by integrating stroke width generated from skeletons of those candidates, we reject remained false positives. Finally, MSERs are clustered into text regions. Experimental results on the ICDAR competition datasets demonstrate that our algorithm performs favorably against several state-of-the-art methods.

Keywords: Performance Evaluation, Historical Document Analysis, Document Understanding
Abstract: Kanungo noise model is widely used to test the robustness of different binary document image analysis methods towards noise. This model only works with binary images while most document images are in grayscale. Because binarizing a document image might degrade its contents and lead to a loss of information, more and more researchers are currently focusing on segmentation-free methods (Angelika et al [2]). Thus, we propose a local noise model for grayscale images. Its main principle is to locally degrade the image in the neighbourhoods of ``seed-points" selected close to the character boundary. These points define the center of ``noise regions". The pixel values inside the noise region are modified by a Gaussian random distribution to make the final result more realistic. While Kanungo noise models scanning artifacts, our model simulates degradations due to the age of the document itself and printing/writing process such as ink splotches, white specks or streaks. It is very easy for users to parameterize and create a set of benchmark databases with an increasing level of noise. These databases will further be used to test the robustness of different grayscale document image analysis methods (i.e. text line segmentation, OCR, handwriting recognition).

Keywords: Graphics Recognition, Layout Analysis, Document Understanding
Abstract: In this paper, we propose a new approach to extract text regions from graphical documents. In our method, we first empirically construct two sequences of learned dictionaries for the text and graphical parts respectively. Then, we compute the sparse representations of all different sizes and non-overlapped document patches in these learned dictionaries. Based on these representations, each patch can be classified into the text or graphic category by comparing its reconstruction errors. Same-sized patches in one category are then merged together to define the corresponding text or graphic layers which are combined to create a final text/graphic layer. Finally, in a post-processing step, text regions are further filtered out by using some learned thresholds.

Keywords: Graphics Recognition, Document Analysis Systems, Document Understanding
Abstract: In this paper, we present an approach for junction detection and reconstruction in line-drawing images. Our approach is shifted to a problem of high curvature point detection from the skeleton of image. Like this, it is independent of any vectorization process and pa- rameter free. The junction reconstruction stage takes benefit of the reliable skeleton segments and their topo- logical relations to reconstruct junctions and make their positions accurate. The experimental results show that the proposed method is competitive with other baseline methods and can achieve accurate junction detection with some pixel errors.

Keywords: Handwriting Recognition
Abstract: HMM is one of the most popular methods to model sequential signals and plays a significant role in the field of off-line handwritten Arabic word recognition research. However, the structure of an HMM including the number of states has to be determined initially and can hardly be updated during the training process. A novel analytic algorithm based on the information entropy of states in an HMM to optimize the number of states will be proposed in this paper. Information entropy is defined as an evaluation criterion of the activity of a state. According to principle of maximum entropy, states with minor information entropy do not possess so enough capability to represent actual observations that they should be deleted. Experiments on IFN/ENIT database show that the algorithm in this paper can bring approximately 3% - 6% increase to correct recognition rate from the best performance of system with constant states.

Keywords: Handwriting Recognition
Abstract: Unconstrained handwritten text recognition systems maximize the combination of two separate probability scores. The first one is the observation probability that indicates how well the returned word sequence matches the input image. The second score is the language probability that reflects how likely a word sequence is according to a language model. Current state-of-the-art recognition systems use statistical language models in form of bi-gram word probabilities. In this paper we propose to model the target language by means of a recurrent neural network with long-short term memory cells. Because the network is recurrent, the considered context is not limited to a fixed size especially as the memory cells are designed to deal with long-term dependencies. In a set of experiments on the IAM off-line database we show the superiority of the proposed language model over statistical n-gram models.

Keywords: Handwriting Recognition, Performance Evaluation
Abstract: In this paper we present a novel segmentation-free Arabic handwriting recognition system based on hidden Markov model (HMM). The proposed system first pre-processes the input image by setting the thickness of the input word to three pixels and fixing the spacing between the different parts of the word. The input image is then decomposed into several images representing the vertical, horizontal, left diagonal and right diagonal edges in the image. A set of robust features representing the densities of the foreground pixels in the various edge images is extracted using sliding windows. The proposed system builds character HMM models and learns word HMM models using embedded training. The performance of the proposed system is very promising compared with other Arabic handwriting recognition systems available in the literature. The contribution of this work is the preprocessing method and the novel feature extraction technique.

Keywords: Handwriting Recognition, Character and Text Recognition, Statistical, Syntactic and Structural Pattern Recognition
Abstract: In this paper, we propose a method to reduce the storage of the incremental MQDF classifier [1] for writer adaptation, through updating the dominant eigenvectors and eigenvalues of MQDF in the incremental discriminant subspace. Compared with previous incremental MQDF algorithm, the proposed Compact Discriminant IMQDF (CDIMQDF) classifier only needs to store the dominant eigenvectors and eigenvalues instead of covariance matrices, thus save the storage cost greatly. The experimental results on SCUT COUCH2009 dataset show that, comparing with previous method, the new CDIMQDF can reduce storage significantly, and in the mean time, keeps comparable or even higher recognition accuracies.

Keywords: Handwriting Recognition
Abstract: In this paper, we propose a novel method for extracting a set of baseline-independent features, which are based on the combination of global and local information. A HMM-based recognition system is developed with 161 models that include a space model and a blank model. All of the models are trained using the standard Baum-Welch Algorithm with the state-tying technique, and are then decoded using the Viterbi Algorithm. Experiments are conducted on the benchmark IFN/ENIT database. Results show that our proposed features can make good use of the relationship between adjacent characters and are sufficiently robust, especially when characters are shifted up or down and when the handwriting width varies.

Keywords: Camera-Based Document Analysis, Character and Text Recognition, Document Analysis Systems
Abstract: To handle the diversity of scene characters, we propose a multiple hypotheses framework which consists of an image operator set module, an optical character recognition (OCR) module, and an integration module. Image operators detect multiple suspicious character areas. The OCR engine is then applied to each detected area and returns multiple candidates with weight values for future integration. Without the aid of heuristic constraints on area, aspect ratio or color etc., the integration module prunes the redundant detection and pads the missing detection based on the outputs of OCR. The experimental results demonstrate that the whole multiple hypotheses outperforms each operator's hypotheses and be comparable with existing methods in terms of recall, precision, F-measure and recognition rate.

Keywords: Document Analysis Systems, Handwriting Recognition, Multimedia Analysis, Indexing and Retrieval
Abstract: To retrieve an arbitrary part of handwritten objects from a database by a handwritten query and be able to deal with subtle variations of object position and shape, the following retrieval method is proposed: (1) Strokes represented by chain codes are classified using a clustering technique. (2) A target object is retrieved using a similarity measure based on an overlap area of circles whose positions and sizes are determined by positions and category types of strokes. We examine the effectiveness of the proposed method using a database with 707 strokes and 30 roughly handwritten queries. As a result, top-10 retrieval accuracy of 97% was obtained.

Keywords: Camera-Based Document Analysis, Detection, Separation and Segmentation, Classification and Clustering
Abstract: Text localization in natural scene images is an important prerequisite for many content-based image analysis tasks. In this paper, we proposed a novel and effective approach to accurately localize scene texts. Firstly, Maximally stable extremal regions(MSER) are extracted as letter candidates. Secondly, after elimination of non-letter candidates by using geometric information, candidate regions are constructed by grouping similar letter candidates using disjoint set. Candidate region features based on horizontal and vertical variances, stroke width, color, stability of MSER and geometry are extracted. An AdaBoost classifier is built from these features and text regions are identified. The overall system is evaluated on the ICDAR 2011 competition dataset and the experimental results show that the proposed algorithm yields high precision and recall compared with the latest published algorithms.

Keywords: Document Analysis Systems, Document Understanding
Abstract: In this paper, a novel content-independent Farsi font recognition algorithm based on the Most-Frequent Connected Components (MFCCs) is proposed. Like previous studies, the font of each text image is uniform. First, all connected components (CCs) of a text are detected. Then, MFCCs of a text image are determined. For each MFCC, some samples are efficiently extracted in the text. After determining the fonts of these samples, the font of the text is recognized by using a voting procedure. Experiments show that the proposed method increases the accuracy of Farsi font recognition.

Keywords: Camera-Based Document Analysis, Stereo and Image-Based Modeling, Document Analysis Systems
Abstract: In this paper we present a novel method for robust stereo matching on document image pairs. The matching itself is performed using an affine-invariant similarity measurement to compensate for perspective distortions, where affine invariance is achieved by normalization using second-order statistics, to finally allow a simple pixel-wise comparison. To handle the inherent high self-similarity of the page content we apply a dynamic programming approach on text-line pairs. We quantitatively show that the proposed method performs better in comparison to standard approaches using SURF - whether with or without incorporating text-line information.

Keywords: Character and Text Recognition, Handwriting Recognition, Performance Evaluation
Abstract: We propose in this work an approach for automatic recognition of printed Arabic text in open vocabulary mode and ultra low resolution (72 dpi). This system is based on Hidden Markov Models using the HTK toolkit. The novelty of our work is in the analysis of three complex fonts presenting strong ligatures: DiwaniLetter, DecoTypeNaskh and DecoTypeThuluth. We propose a feature extraction based on statistical and structural primitives allowing a robust description of the different morphological variability of the considered fonts. The system is benchmarked on the Arabic Printed Text Image (APTI) database.

Keywords: 2D/3D Object Detection and Recognition, Features and Image Descriptors, Segmentation, Color and Texture
Abstract: Many three-dimensional (3-D) image-based studies concerning granular and sintered materials require a description of the observed microstructures in terms of individual grains. We propose a robust segmentation algorithm which identify groove regions on the object�s surface in order to locate possible grain boundaries in the object�s volume. The algorithm relies on the volumetric propagation via Voronoi labeling of curvature information from the surface into the object

Keywords: 2D/3D Object Detection and Recognition, Detection, Separation and Segmentation, Image and Video Understanding
Abstract: This paper describes an efficient acceleration of GAT (Global Affine Transformation) correlation as a powerful technique of distortion-tolerant image matching. The key ideas are twofold: efficient calculation of optimal affine parameters that maximize the normalized cross-correlation value between an input image and a template via separation of variables in the original GAT computational model and generation of lookup tables to further reduce the computational cost. Experiments using a template and its artificially affine-transformed images as input images demonstrate that the accelerated GAT correlation technique can absorb a fair amount of affine transformation much accurately than the wellknown tangent distance method. Furthermore, the computation time of the accelerated GAT correlation is greatly reduced to less than five percent of that of the original GAT correlation. As a result, the processing time of the accelerated GAT correlation is around a half of that of the tangent distance method.

Keywords: Features and Image Descriptors, 2D/3D Object Detection and Recognition, Vision for Robotics
Abstract: In this paper, we present a new method for a locally adaptive region detector called Bilateral kernel-based Region Detector (BIRD). This work is to detect stable regions from images by consecutively computing a multiscale decomposition based on the bilateral kernel. The BIRD regards a region as covariant if it exhibits predictability in its photometric distance over spatial distance. Distinctiveness and robustness across scales are achieved by selecting the extremely stable regions through sequential scales. Our method is simple and easy to implement. Experimental results show that our method outperforms competing affine region detection methods in efficiency on region detection.

Keywords: Motion, Tracking and Video Analysis
Abstract: The Horn-Schunck (HS) optical flow method is widely employed to initialize many motion estimation algorithms. In this work, a variational Bayesian approach of the HS method is presented where the motion vectors are considered to be spatially varying Student's t-distributed unobserved random variables and the only observation available is the temporal image difference. The proposed model takes into account the residual resulting from the linearization of the brightness constancy constraint by Taylor series approximation, which is also assumed to be a spatially varying Student's t-distributed observation noise. To infer the model variables and parameters we recur to variational inference methodology leading to an expectation-maximization (EM) framework in a principled probabilistic framework where all of the model parameters are estimated automatically from the data.

Keywords: Motion, Tracking and Video Analysis, 2D/3D Object Detection and Recognition, Scene Understanding
Abstract: Given a single image of a scene rectangle of an unknown aspect ratio and size, we present a method to reconstruct the projective structure and to find camera parameters including focal length, position, and orientation. First, we solve the special case when the center of a scene rectangle is projected to the image center. We formulate this problem with coupled line cameras and present the analytic solution for it. Then, by prefixing a simple preprocessing step, we solve the general case without the centering constraint. We also provides a determinant to tell if an image quadrilateral is a projection of a rectangle. We demonstrate the performance of the proposed method with synthetic and real data.

Keywords: Segmentation, Color and Texture, Detection, Separation and Segmentation, Image and Video Processing
Abstract: Graph-based segmentation requires interactive input, and color segmentation poses challenges when foreground has similar color with background. This paper presents an automatic robust segmentation fusing color and depth. A saliency detection algorithm based on visual attention is proposed. Depth is firstly used as a filter to limit object pixel candidates, color is used to further detect object region by saliency calculation. The saliency detection algorithm is incorporated into graph-based segmentation by automatically setting the detected range as uncertain pixel set. Our segmentation algorithm uses color as main information to set energy terms of graph function, and uses depth as supplemental information to adjust these terms basing on consistency decision of color and depth. Experimental results demonstrate that the proposed segmentation enhances the performance using color alone, and realizes fully automation.

Keywords: Segmentation, Color and Texture, 2D/3D Object Detection and Recognition
Abstract: In this paper, we address the problem of representing objects using contours for the purpose of recognition. We propose a novel segmentation method for integrating a new contour matching energy into level set based segmentation schemes. The contour matching energy is represented by major components of Elliptic Fourier shape descriptors and serves as a shape prior to guide the curve evolution. The contours in training dataset serve as templates and are utilized to infer the category of an unknown image based on matching. Our method is evaluated on the UCF sports dataset and Caltech 101 dataset. Experiments show that our method achieves promising recognition accuracy and is robust to noisy low-level features and background clutter.

Keywords: Features and Image Descriptors, Machine Learning and Data Mining, Classification and Clustering
Abstract: In this paper, we propose an unsupervised cluster method via a multi-task learning strategy, called textbf{MtCluster}. Our MtCluster learns a cluster-specific dictionary for each cluster to represent its sample signals and a shared common pattern pool (the commonality) for the essentially complemental representation. By treating learning the cluster-specific dictionary as a single task, MtCluster works in a multi-task learning manner, in which all the tasks are connected by simultaneously learning the commonality. Actually, the learned cluster-specific dictionary spans the feature space of the corresponding cluster, and the commonality is just used for necessary complemental representation. To evaluate our method, we perform several experiments on public available datasets, and the promising results demonstrate the effectiveness of MtCluster.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Surveillance and Security, Image and Video Processing
Abstract: Under-water video-mosaics are an important tool e.g. for inspection of man-made infrastructure. Cameras may drift in rotation and distance to the surface while the mosaic will often be very much larger than a single frame resulting in long chains of planar homographies. This contribution addresses the problems arising from dead-reckoning drift in such chains. Local patches are rectified using homography decomposition. Experiments with a portal allowing underwater image sequences with mechanical groundtruth are performed. Thus the deviation of the mosaic from a true orthographic map may be estimated.

Keywords: 2D/3D Object Detection and Recognition, Features and Image Descriptors, Scene Understanding
Abstract: Intuitive and easily interpretable performance measures, repeatability and matching performance, for local feature detectors and descriptors were introduced by Mikolajczyk et al. (IJCV2005,PAMI2005). They, however, measured performance in a wide baseline setting that does not correspond to the visual object categorisation problem which is a popular application of the detectors and descriptors. The limitation has been recognised and ad~hoc evaluations proposed. To the authors' best knowledge, our work is the first which extends the original repeatability and matching performance measures to the case of object classes. Using the novel evaluation framework we test state-of-the-art detectors and descriptors with the popular Caltech-101 dataset and report the object category level (intra-class) repeatability and matching performances.

Keywords: Motion, Tracking and Video Analysis, Image and Video Understanding, Pattern Recognition for Surveillance and Security
Abstract: This paper proposes a method to estimate earthquake ground motion by analyzing the video taken with a fixed surveillance camera. In recent years, a network of seismometers was constructed to observe the ground motion. However, the spatial density of the seismometers was not high enough to obtain an adequate spatial distribution of the earthquake ground motion, which varies depending on the local ground condition across tens or hundreds of meters. To interpolate the network of seismometers, we propose a method to estimate earthquake ground motion by using a fixed surveillance camera. In our method, the motion of objects sliding on the ground or on a floor is estimated by image analysis. Then, the earthquake ground motion is obtained by inverse analysis of the sliding motion. We conducted scale model experiments to evaluate our method, and the experimental results show the effectiveness of our method.

Keywords: Stereo and Image-Based Modeling
Abstract: By taking advantage of recent progress in patch match based stereo, we propose a depth-map merging based Multi-View Stereo (MVS) reconstruction method for large-scale scenes . Combining the strength of the accuracy of patch match stereo and the redundancy information provided by multiple views, our method can produce quite satisfactory MVS reconstruction results. Besides, the proposed method could be easily parallelized at image level, i.e., each depth-map is computed individually, which makes it suitable for large-scale scene reconstruction with high resolution images.

Keywords: Segmentation, Color and Texture, Detection, Separation and Segmentation, Image and Video Processing
Abstract: We introduce the vectorial Minimum Barrier Distance (MBD), a method for computing a gray-weighted distance transform while also incorporating information from vectorial data. Compared to other similar tools that use vectorial data, the proposed method requires no training and does not assume having only one background class. We describe a region-growing algorithm for computing the vectorial MBD efficiently. The method is evaluated on two types of multichannel images: color images and textural features. Different path-cost functions for calculating the multidimensional path-cost distance are also compared. The results show that by combining multi-channel images into vectorial information the performance of the vectorial MBD segmentation is improved compared to when one channel is used. This implies that the method can be a good way of incorporating multichannel information in interactive segmentation.

Keywords: Motion, Tracking and Video Analysis, Vision for Robotics
Abstract: In this paper, we present a novel method for stereo camera-based 3D tracking which integrates point-kinematics, associated to each visual feature into Kalman filters. The approach utilizes optical flow and stereo correspondence of visible, predominatly specular features on a target satellite surface,in order to estimate translational and rotational velocities of the rigid body. The motion of each 3D point cloud can be predicted, since all point clouds are constrained by the common, rigid motion. A dual quaternion-based pose estimator, robustified with median statistics, is further applied to the estimated points. In case of temporarily missing measurements, the last estimated body velocity is used to predict the next poses.Experimental results based on images of a satellite simulator are shown to demonstrate performances for on-orbit servicing.

Keywords: Segmentation, Color and Texture, Low-Level Vision, 2D/3D Object Detection and Recognition
Abstract: State-of-the-art snake methods for object segmentation fail in the presence of strong clutter. Here, we present a dynamic programming (DP) setup to combat strong clutter. Our solution maximizes a score function known as gradient inverse coefficient of variation (GICOV). GICOV cannot be directly used in DP, because it does not have an additive form. We derive a set of DP-friendly necessary conditions for maximization of the GICOV. Our experiments illustrate that while other snake methods are thwarted by clutter, the proposed method finds particle object boundaries rejecting clutter and distracters.

Keywords: Segmentation, Color and Texture, Inpainting and Superimposing, Vision for Robotics
Abstract: Accurate segmentation of sidewalks from satellite images can be required in various applications, for example giving walking directions to pedestrians and robot navigation. We propose a framework to construct sidewalk and crosswalk maps from satellite images. This is a challenging task, since typically sidewalks in satellite images are highly occluded by trees and their shadows and also there can be several objects on maps that have similar appearance with sidewalks. In our framework, we initially segment visible sidewalks by their appearance and then complete the tree-occluded areas by posing the problem as an image-inpainting task with multiple priors, which fuses the knowledge about roads, occluders and sidewalk structures. We present successful sidewalk segmentation results from satellite images, where sidewalks are highly occluded by trees.

Keywords: 2D/3D Object Detection and Recognition, Image and Video Processing, Pattern Recognition for Bioinformatics
Abstract: Varying lighting conditions may cause great face image appearance variance and hence seriously distort the distributions of within and between class distances, but these distributions are exactly the foundation of successful automatic recognition. In this paper, a dynamic illumination distinction description strategy is put forward to compensate for the lighting difference between probe and gallery faces. After enrolling this description, face image distances will more faithfully reflect face identity difference. We also introduce an algorithm framework to integrate both illumination distinction description computation and face recognition into once l1-optimization. The effectiveness of proposed approach is verified on the Extended Yale B dataset.

Keywords: 2D/3D Object Detection and Recognition
Abstract: We propose a high-accuracy human detection method featuring a Haar-like filter expressing the human shape and using depth information obtained by capturing people from above with a time-of-flight (TOF) camera. This method extracts object regions by performing background subtraction against this depth information, and passes these extracted object regions through a Haar-like filter based on a human model expressing the convex shape of shoulder-head-shoulder. Human detection is achieved by integrating the results of this filtering by mean-shift clustering. The proposed method improves detection rate by 5.7% compared to a human-detection technique that simply applies mean-shift clustering to depth information obtained by background subtraction. We show that our method can detect humans in real time at a frame rate of about 19 fps.

Keywords: Motion, Tracking and Video Analysis, Image and Video Understanding, 2D/3D Object Detection and Recognition
Abstract: The aim of this paper is to track objects during their use by humans. The task is difficult because these objects are small, fast-moving and often occluded by the user. We present a novel solution based on cascade action recognition, a learned mapping between body- and object-poses, and a hierarchical extension of importance sampling. During tracking, body pose estimates from a Kinect sensor are classified between action classes by a Support Vector Machine and converted to discriminative object pose hypotheses using a {body, object} pose mapping. They are then mixed with generative hypotheses by the importance sampler and evaluated against the image. The approach out-performs a state of the art adaptive tracker for localisation of 14/15 test implements and additionally gives object classifications and 3D object pose estimates.

Keywords: 2D/3D Object Detection and Recognition, Vision for Robotics
Abstract: In this paper we propose a generic 6d object localization approach based on surface normal images and CAD model data. Normal images or �normal maps� can be obtained using only one single camera shot of a simple camera-projector system. The advantages of this sensor setup are very short acquisition times and the exclusive use of consumer hardware, namely a projector and a grey value camera, making it easy and cheap to build up. All computations are analytically tractable and can be computed in the Fourier domain allowing a very effective and fast computation.

Keywords: Scene Understanding, Pattern Recognition for Search, Retrieval and Visualization, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Due to the well-known semantic gap, a new trend is to apply conceptual hierarchies to improve annotation performance. Benefiting from the increasing availability of user contributed images with predefined groups and tags, we present a novel group-based probabilistic latent semantic analysis (PLSA) model for automated image annotation. For each group, we assume that the latent topics are mapped in a sub-manifold of the ambient space and develop a new EM algorithm for the group-based PLSA model. Then, we annotate an image with group mapping and tag allocation. Finally, comparing with several state-of-the-art approaches on Corel5K and Flickr datasets, our approach shows a promising performance.

Keywords: Motion, Tracking and Video Analysis, Human Computer Interaction, Gesture and Behavior Analysis
Abstract: This paper introduces a new vision-based real-time motion capture system for the applications in consumer electronics. The motion capturing systems face two challenges: parameter estimation in high dimension space and self-occlusion. This system uses the locality sensitive hashing (LSH) to find the similar examples and then estimates the pose parameters. The contributions of this example-based method are (1) proposing the modified LSH function for fast example patches retrieval, (2) applying Hough voting to estimate the pose parameters, and (3) adding the temporal constraints and the prediction constraints before and after Hough voting to increase the prediction accuracy.

Keywords: Motion, Tracking and Video Analysis
Abstract: Online learning has shown to be powerful on tracking of previously unknown objects. However, tracking non-rigid objects is still challenging. In this paper, an object tracking method based on modified online Random Naive Bayes and GrabCut segmentation is proposed. Firstly, the online Random Naive Bayes integrated with Hough voting is used to identify the foreground and background pixels. Then, the classification result is refined with the GrabCut segmentation. Finally, the training samples are relabeled according to the segmentation regions, which can significantly reduce the amount of wrong labeled samples during online learning and against drifting problem. The experimental results show that the proposed scheme is robust in tracking objects in challenging sequences.

Keywords: Vision for Robotics, Mixed and Augmented Reality
Abstract: Existing planar visual markers (AR markers) have poor accuracy in pose estimation, especially in frontal direction. We solved the problem by a novel principle using a microlens array. The marker displays a changing two-dimensional moire pattern according to the visual-line direction. We can extract pose information from the pattern by image processing. The developed marker and the processing algorithm enable high-accuracy pose estimation even by observation from frontal direction. We verified its superiority to the conventional method by some validation tests.

Keywords: Motion, Tracking and Video Analysis, 2D/3D Object Detection and Recognition
Abstract: In this paper, we address the problem of reconstructing 3D trajectories given only 2D point projection trajectories of an articulated structure. Hitherto, most applications of articulated trajectory reconstruction require: (i) relative lengths, and (ii) camera motion. We propose a novel extension that allows us to in many circumstances circumvent these limitations. Of particular note is our characterisation theoretically and demonstration empirically how agnostic one be to camera motion when using a DCT basis. Practical reconstructions of non-human articulated structures (e.g. koala, sea snake, tasmanian tiger) where articulated lengths and camera motion are unknown.

Keywords: Vision for Graphics, Features and Image Descriptors, 2D/3D Object Detection and Recognition
Abstract: The detection of salient regions from mesh surfaces is an important preprocessing step for many 3D appli- cations, such as mesh simplification, registration, seg- mentation and compression, etc. The detected salient regions facilitate the understanding of the structure and finding the regions/components that are important on 3D surfaces. This paper proposes a novel method for saliency detection by diffusing the shape index field with non-local means filter, generating a saliency map, that further enhanced using Retinex approach. The ef- fectiveness of this method is demonstrated by surface simplification. Experimental results demonstrate that the proposed approach has achieved better results than competitors.

Keywords: Features and Image Descriptors, Pattern Recognition for Search, Retrieval and Visualization, Multimedia Analysis, Indexing and Retrieval
Abstract: The conventional EM algorithms may suffer from the following two problems. First, it may converge to a local maximum. Second, the algorithm may suffer from singularity. A novel Enhanced EM algorithm (EEM) using a realization of maximum-entropy uniform distribution as initial condition is proposed. A global optimal solution can be obtained. In addition, a positive perturbation scheme is adopted to avoid singularity. Experimental results have demonstrated that the EEM is simple and effective compared with some prior arts.

Keywords: Features and Image Descriptors, Low-Level Vision, Detection, Separation and Segmentation
Abstract: A model-based method for transformation-invariant area descriptor extraction is proposed in this paper in the context of object recognition and image matching. Local image descriptors are extracted in salient circular fragments of variable size, which indicate image locations with high intensity contrast, regional homogeneity and shape saliency. Three different types of descriptors � pose, intensity, and area shape � are extracted to form a single descriptor vector. The pose and intensity descriptors are made relational and normalized in order to achieve invariance to image similarity transformations and affine intensity changes. The method requires no image segmentation since the feature points are time-efficiently extracted in a multi-scale manner by analyzing circular areas of various sizes.

Keywords: Vision for Robotics, Scene Understanding, Physics-Based Vision
Abstract: It has been shown, in previous work, that the 3D position of a line can be reconstructed from a single image in vision systems that do not possess a single viewpoint. We present a new method that, in a non-central axial catadioptric system, can achieve line spacial reconstruction from 3 or more image points, given the distance ratio of 3 points in the line (a fair assumption in, for example, structured environment with repetitive architectural features). We use cross-ratio as an invariant to constrain the line localization and perform the reconstruction from a set of image points through non-linear optimization. Experimental results are presented.

Keywords: Motion, Tracking and Video Analysis, Detection, Separation and Segmentation, Segmentation, Color and Texture
Abstract: This paper introduces an algorithm for dense motion segmentation using an �articulated motion invariant distance� in crowded video sequences. This algorithm solves two conventional problems: one is dense motion segmentation of pedestrians with articulated motion; the other is severe occlusions especially in a crowded scene. Unlike existing methods, the proposed algorithm uses temporal invariance of geodesic distance (similarity) between segments as a criterion for motion segmentation. It can segment the whole appearance of each articulated object, while articulated object contains a wide variety of motions even on the same object. To solve the �severe occlusion problem�, a random walk framework is applied to Intersegmental Spatio-Temporal Connectivity (ISTC) to obtain dense and long-term temporally-coherent segments from short-term temporally-incoherent segments. The experimental results show that the algorithm can successfully obtain dense and temporally coherent segments under severe occlusion conditions.

Keywords: Motion, Tracking and Video Analysis
Abstract: This paper proposes an activity-specific 3D human pose tracking system from multiple camera views. Dimensionality reduction is used to represent a single activity in a hierarchy of low dimensional spaces. This hierarchy provides increasing independence between limbs by decoupling them, allowing higher flexibility and adaptability that result in improved accuracy. For every subspace, a deterministic optimisation method is applied to estimate the position of the corresponding body parts. Searching through the hierarchy is controlled by an observation function to minimise the computational cost. Evaluation on HumanEva sequences demonstrates that the proposed framework is state-of-the-art both in terms of accuracy and computational complexity.

Keywords: Motion, Tracking and Video Analysis, Low-Level Vision, Image and Video Processing
Abstract: In this work, we introduce an alternative measurement to evaluate quality of flow field. We then present an alternative filter called OA-filter (Occlusion-Aware filter) which uses this measurement. It is integrated inside the solver to push the performance of the model in terms of accuracy and sharpness. The experiments show that the estimation result has been significantly improved. The method yields sharp flow field which is very useful for video processing, 3D reconstruction, video tracking, and moving object detection.

Keywords: Motion, Tracking and Video Analysis, Scene Understanding
Abstract: This paper proposes a generic approach combining a bottom-up (low-level) visual detector with a top-down (high-level) fuzzy first-order logic (FOL) reasoning framework in order to detect pedestrians from a moving vehicle. Detections from the low-level visual corner based detector are fed into the logical reasoning framework as logical facts. A set of FOL clauses utilizing fuzzy predicates with piecewise linear continuous membership functions associates a fuzzy confidence (a degree-of-truth) to each detector input. Detections associated with lower confidence functions are deemed as false positives and blanked out, thus adding top-down constraints based on global logical consistency of detections. We employ a state of the art visual detector on a challenging pedestrian detection dataset, and demonstrate an increase in detection performance when used in a framework that combines bottom-up detections with (fuzzy FOL-based) top-down constraints.

Keywords: Segmentation, Color and Texture, Classification and Clustering, Detection, Separation and Segmentation
Abstract: In this paper, we will propose a fast K-means algorithm based on knowledge learning (FKA), and apply it for real-time image segmentation. Compared with traditional K-means clustering algorithms, FKA focuses on how to make use of knowledge in the process of clustering to speed up the convergence of traditional K-means algorithm. Specifically, FKA accelerates the clustering process by the information obtained in the previous iterations. The performance of FKA is analyzed theoretically. The experiments show that FKA outperforms most of the state-of-the-art K-means clustering algorithms when applied to large image segmentation.

Keywords: Low-Level Vision, Enhancement, Restoration and Filtering, Image and Video Processing
Abstract: In this paper, a novel sparse representation based super-resolution (SR) method is proposed to reconstruct a high resolution (HR) face image from a low resolution (LR) observation via training samples. First, a specific LR and HR over-complete dictionary pair is learned for a certain patch over the patches in all training samples with the same position. Second, K-selection mean constrain is used to make the sparse representation of the input patch more accurate. Third, the HR patch is hallucinated via the sparse representation coefficients and the HR dictionary. At last, we form the final HR face image by integrating the hallucinated HR patches together. Experiments validate the proposed method in extensive data. Compared to some state-of-the-art methods, our method exhibits better performance both in subjective and objective quality.

Keywords: 2D/3D Object Detection and Recognition, Motion, Tracking and Video Analysis, Image and Video Processing
Abstract: For safety purpose, railroad tracks must be inspected regularly for defects or other design non-compliances. One crucial building block in an automatic inspection system is to detect different types of railroad track objects. We introduce a novel global optimization framework to combine evidence from multiple cameras and the distance measuring instrument to improve rail object detection. Our framework leverages the cross-object spatial constraints enforced by the sequential structure of rail tracks, as well as the cross-frame and cross-view constraints in camera streams. Experimental results on real rail track-driving data demonstrates that our approach achieves superior performance compared to processing each data stream independently. We argue that our approach can be extended to other embodiments involving linear sequential structures, such as pipeline, highway and road inspection.

Keywords: Segmentation, Color and Texture, Detection, Separation and Segmentation
Abstract: We propose a method for extracting a desired object in multi-view images without camera calibration. We match the corner points obtained automatically by the Scale-invariant feature transform (SIFT) in multi-view images, and then connect multi-view images into a weighted undirected graph. Thus, multi-view object segmentation converts to a graph partitioning problem solved by Biased Normalized Cuts. Compared to the existing methods, the main advantages of our method are that: (1) it needn't camera pose and intrinsic parameters; (2) arbitrary number images (include single image) can be processed. The promising experimental results for images reveal the effectiveness of our approach.

Keywords: Segmentation, Color and Texture, Image and Video Processing
Abstract: Otsu's method is considered as one of the most popular thresholding method due to its simplicity. In order to increase its capabilities, one approach is to additionally use multidimensional information so that noise in images can be handled; while the other approach is to allow multilevel instead of bilevel to extract objects in the case of complex images. These two approaches distinctly offer different advantages, where real world images mostly contain both kinds of image difficulties. In this paper, we propose the use of both multidimensional information and multilevel thresholding together, so that we can efficiently cope with real world images. From the experiments, even though our method requires a bit longer execution time, the method is more resist to noise and also give better or comparable results than the others.

Keywords: Motion, Tracking and Video Analysis
Abstract: Recently, fusion of low- and high-dimensional approach shows its success in the generic human motion tracking. However, how to choose the tracker adaptively according to the motion types is still a challenging problem. This paper presents a trackers sampling approach for generic human motion tracking using both low- and high-dimensional trackers. Gaussian Process Dynamical Model(GPDM) is trained to learn the motion model of low-dimensional tracker, and it performs better on specific motion types. Annealed Particle Filtering( APF) shows its advantage in the tracking without limitation on motion types. We combine both of the two methods and automatically sample trackers according to the motion types that it is tracking on. To improve performance, trackers communication is adopt to keep the better state of trackers. The approach facilitates tracking of genetic motions with low particle numbers.

Keywords: Scene Understanding, Low-Level Vision
Abstract: In this paper we propose a novel method of estimating vanishing point by spherical gradient. In contrast with the conventional methods in which vanishing point is estimated from lines, the proposed method does not necessarily extract lines explicitly, but employs the spherical gradient of edge points. Based on the observation that spherical gradient is aligned with the normal vector of the projection plane of space lines, the vanishing point is estimated directly from spherical gradient of edge points by the Hough Transform.

Keywords: Motion, Tracking and Video Analysis
Abstract: To deal with the drifting issue in visual tracking, we propose an Online Transfer Boosting (OTB) algorithm that transfers knowledge from three different source domains to the target domain to improve the performance of the online classifier used in tracking-by-detection. In particular, the OTB algorithm integrates three types of knowledge by: (1) transferring prior knowledge from the first frame using semi-supervised learning; (2) transferring appearance changes from the previous frames by dynamically updating the learning factor; and (3) transferring observed sample distribution knowledge from the current frame by reweighting the training samples. Experimental results on several public video sequences demonstrated promising performance of OTB in both tracking accuracy and stability.

Keywords: 2D/3D Object Detection and Recognition
Abstract: This paper addresses the problem of Car Make and Model recognition as an example of within-category object class recognition. In this problem, it is assumed that the general category of the object is given and the goal is to recognize the object class within the same category. As compared to general object recognition, this problem is more challenging because the variations among classes within the same category are subtle, mostly dominated by the category overall characteristics, and easily missed due to pose and illumination variations. Therefore, this specific problem may not be effectively addressed using generic object recognition approaches. In this paper, we propose a new approach to address this specific problem by combining global and local information and utilizing discriminative information labeled by a human expert. We validate our approach through experiments on recognizing the make and model of sedan cars from single view images.

Keywords: Stereo and Image-Based Modeling, Mixed and Augmented Reality, Image and Video Processing
Abstract: This paper proposes a new cost construction method with multiscale Weber (MSW) descriptor and weighted linear regression for robust stereo matching in a two-layer hierarchical structure. Firstly, the MSW descriptors extracted from stereo pairs are utilized to combined raw matching costs to reduce the disparity search range. Secondly, the indispensable matching costs on the subsets of disparity candidates are aggregated by spatial moving discrete sampling and weighted least square. The proposed cost technique can provide an effective and efficient way to reflect robust visual dissimilarity measure for local stereo approach. Experimental results on the Middlebury benchmark datasets demonstrate the validity of the presented cost construction which can eliminate effectively the ambiguity of stereo matching.

Keywords: Features and Image Descriptors, Low-Level Vision
Abstract: Selectivity and invariance are thought to be important ingredients in biological or artificial visual systems. A fundamental problem is, however, to know what the visual system should be selective to and what to be invariant to. Building a statistical model of images, we learn here a three-layer feature extraction system where the selectivity and invariance emerges from the properties of the images.

Keywords: Segmentation, Color and Texture, Detection, Separation and Segmentation, Image and Video Understanding
Abstract: Semantic image segmentation assigns a predefined class label to each pixel. This paper proposes a unified framework by using region bank to solve this task. Images are hierarchically segmented leading to region banks. Local features and high-level descriptors are extracted on each region of the bank. Discriminative classifiers are learned based on the histograms of feature descriptors computed from training region bank (TRB). Optimally merging predicted regions of query region bank (QRB) results in semantic labeling. Each algorithmic module used in our system is detailed, and as the proposed framework is generic, any algorithm which fits corresponding modules can be plugged into the framework. Experiments on the challenging Microsoft Research Cambridge (MSRC 21) dataset show that the proposed approach achieves the state-of-the-art performance.

Keywords:
Abstract: In this talk, we will discuss a new class of models and techniques that can effectively model and extract rich low-dimensional structures in high-dimensional data such as images and videos, despite nonlinear transformation, gross corruption, or severely compressed measurements. This work leverages recent advancements in convex optimization for recovering low-rank or sparse signals that provide both strong theoretical guarantees and efficient and scalable algorithms for solving such high-dimensional combinatorial problems. These results and tools actually generalize to a large family of low-complexity structures whose associated (convex) regularizers are decomposable. We illustrate how these new mathematical models and tools could bring disruptive changes to solutions to many challenging tasks in computer vision, image processing, and pattern recognition. We will also illustrate some emerging applications of these tools to other data types such as web documents, image tags, microarray data, audio/music analysis, and graphical models. This is joint work with John Wright of Columbia, Emmanuel Candes of Stanford, Zhouchen Lin of Peking University, and my students Zhengdong Zhang, Xiao Liang of Tsinghua University, Arvind Ganesh, Zihan Zhou, Kerui Min and Hossein Mobahi of UIUC.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Surveillance and Security, Classification and Clustering
Abstract: In this paper, we focus on a challenging pattern recognition problem of significant industrial impact: classifying vehicles from their rear videos as observed by a camera mounted on top of a highway with vehicles traveling at high speed. To solve this problem, we present a novel feature called structural signatures. From a rear view video, a structural signature recovers the vehicle side profile information which is crucial in its classification. As a vehicle moves away from a camera, its surfaces deform differently based on their relative orientation to the camera. This information is used to extract the structure of the vehicle which captures the relative orientation of vehicle surfaces and the road surface. We present a complete system which computes the structural signatures and uses them for classification of passenger vehicles into sedans, pickups and Minivans/SUVs in highway videos. We analyze performance of the system on a large dataset.

Keywords: Segmentation, Color and Texture
Abstract: Superpixel segmentation has become a popular preprocessing step in computer vision with a great variety of existing algorithms. Almost all algorithms claim to compute compact superpixels, but no one showed how to measure compactness and no one investigated the implications. In this paper, we propose a novel metric to measure superpixel compactness. With this metric, we show that there is a trade-off between compactness and boundary recall. In addition, we propose an algorithm that allows to precicely control this trade-off and that outperforms the current state-of-the-art. As a demonstration, we show the importance of considering compactness with the help of an example application.

Keywords: Features and Image Descriptors, Low-Level Vision
Abstract: Local Binary Descriptors (LBDs) are good at matching image parts, but how much information is actually carried? Surprisingly, this question is usually ignored and replaced by a comparison of matching performances.

In this paper, we directly address it by trying to reconstruct plausible images from different LBDs such as BRIEF and FREAK. Using an inverse problem framework, we show that this task is achievable with only the information in the descriptors, excluding the need of additional data. Hence, our results represent a novel justification for the performance of LBDs.

Furthermore, since plausible images can be inferred using only these simple measurements, this emphasizes the concerns about privacy and secrecy of image keypoints raised previously by Weinzaepfel et al., that could have an important impact on public applications of image matching.

Keywords: Coding and Compression, Image and Video Processing
Abstract: As an extension of AVC, SVC provides the ability to adapt to heterogeneous environments. However, transcoding between SVC and AVC becomes necessary due to the existence of legacy AVC-based systems. This paper proposes a low-complexity SVC-to-AVC CGS transcoder in the pixel domain, which achieves approximately the same coding efficiency as the full re-encoding method. The output AVC bitstream modes are deduced by the proposed mode-mapping strategy. Different from previous works which simply use full search for SVC modes, the SVC modes are also examined and refined to generate suitable AVC modes in the proposed method. Simulation results show that proposed transcoder achieves 10 times speed-up with only 2% BDBR increase and 0.08 dB BDPSNR loss, compared with the full re-encoding method.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing, Segmentation, Color and Texture
Abstract: This paper proposes an approach whereby the regional contrasts in snap-shot style portrait photographs are enhanced using pre-modern portrait paintings as aesthetic exemplars. The reference portrait painting is selected based on a comparison of the existing contrast properties of the painting and the photograph. The contrast style in the selected reference painting is transferred to the photograph by mapping the inter- and intra-regional contrast of the regions, such as the face and skin area, reminder part of figure, and non-figure region. A novel two-part partial contrast stretching method is proposed to achieve the contrast mapping. Finally, the transition boundary between regions is smoothed and the non-figure region is blurred. The experiment results demonstrate the effectiveness of the proposed method.

Keywords: Image and Video Processing
Abstract: In this paper, based on a novel integer transform, an efficient reversible image watermarking method is proposed. The integer transform is a generalization of some previous works. It can be applied to a pixel block of arbitrary size with adjustable capacities. Moreover, this transform is conducted by conditionally embedding data only when the estimated distortion is acceptable. By this approach, our method can effectively reduce the embedding distortion while providing a high capacity. The experimental results show that the proposed method outperforms some state-of-the-art works.

Keywords: Coding and Compression, Image and Video Processing
Abstract: This paper proposes a motion compensation method using sparse representation for video coding technique. A new generation video compression technology known as the HEVC (High Efficiency Video Coding) suggests a novel motion compensation called local intensity compensation. Local intensity compensation represents a current block by a linear combination of some reference blocks. In this study, the weight coefficients of the linear combination are calculated by sparse representation. Sparse representation is able to obtain many zero coefficients, thus the current block can be represented by a few and meaningful reference blocks. The experimental observations demonstrate that the quality of local intensity compensation increases by sparse representation.

Keywords: Coding and Compression
Abstract: With the recent explosion in the development of multimedia hardware capable of 3D display, 3D Picture Coding Sytems have assumed a pivotal role. While encoding techniques for stereo-scopic images is a well researched topic and compression standards such as MPEG provide variants to support it, compression of RGB-D data such as from the Microsoft Kinect sensor offers a number of unsolved challenges. Projected texture based active sensors such as the Kinect offer a number of advantages in comparison with traditional 3D capture systems. While not affecting the visible spectrum of the scene these sensors are capable of producing highly accurate 3D reconstructions of complex scenes (as well as novel viewpoints) - even those with homogeneous surfaces lacking textural features that form the foundation of stereoscopic range measurement systems. Conventional approaches to compressing the RGB and D images separately are suboptimal in terms of compression efficiency, bandwidth usage and scalability. On the other hand, state-of-the-art methods in the field are not suitable for low bandwidth applications, typical of mobile phone devices, on-field civilian and defense robotic systems, especially those operating on unreliable or high-loss wireless networks. In order to address these concerns, we present a novel RGB-D Cross-Compression algorithm that can be used for static 3D scene reconstruction as well as intra-frame coding of 3D videos. The algorithm detects salient edge-like structures in RGB and D images and perform cross-coding across the modalities to yield a scalable system for 3D video coding. Results presented using the Microsoft Ballet and Breakdancers test sequences demonstrate the efficiency of the system in terms of reconstruction quality and rate-distort characteristics.

Keywords: Feature Reduction and Manifold Learning
Abstract: Recently, L1 graph based analysis using sparse representation has received a lot of attentions in pattern recognition and related communities. In this paper, motivated by the success of L1 graph in dimensionality reduction, we extend it to feature selection and propose a novel filter-type method called Sparsity Score (SS) which ranks features according to their respective sparsity preserving capability. For that aim, a L1 graph is constructed based on sparse representation on samples, where a L1-norm based optimization is used to simultaneously determine the graph adjacency structure and corresponding graph weights of the L1 graph. Experimental results on a series of benchmark data sets show that the proposed SS method achieves better performances than conventional feature selection methods.

Keywords: Machine Learning and Data Mining, Feature Reduction and Manifold Learning, Classification and Clustering
Abstract: Nonnegative Matrix Factorization (NMF), a popular compact data representation method, fails to discover the intrinsic geometrical structure of the data space. Graph regularized NMF (GrNMF) is proposed to avoid this limitation by regularizing NMF with a nearest neighbor graph constructed from the input data feature space. However using the original feature space directly is not appropriate because of the noisy and irrelevant features. In this paper, we propose a novel data representation algorithm by integrating feature selection and graph regularization for NMF. Instead of using a fixed graph as GrNMF, we regularize NMF with an adaptive graph constructed according to the feature selection results. A uniform object is built to consider feature selection, NMF and adaptive graph regularization jointly, and a novel algorithm is developed to update the graph, feature weights and factorization parameters iteratively. Data clustering experiment shows the efficacy of the proposed method on the Yale database.

Keywords: Machine Learning and Data Mining
Abstract: In many real-world applications, different mis-classification errors will cause different costs. However, cost-sensitive learning only applied in classification phase and not in the feature selection phase to address this problem. In this paper, we study cost-sensitive feature selection and propose a framework which incorporates a cost matrix into traditional feature selection methods. And we developed three corresponding methods, namely, Cost-Sensitive Variance Score (CSVS), Cost-Sensitive Laplacian Score (CSLS), Cost-Sensitive Constraint Score (CSCS). Experiments on real software defect prediction benchmark data sets demonstrate that cost-sensitive feature selection methods are more efficacy than traditional ones in reducing the total cost.

Keywords: Feature Reduction and Manifold Learning, Features and Image Descriptors, Statistical, Syntactic and Structural Pattern Recognition
Abstract: This paper proposes a framework of tensor-based ICA method for N-dimensional data analysis, which is called generalized N-dimensional ICA (GND-ICA). The proposed GND-ICA is based on multilinear algebra that treats N-dimensional data as a tensor without any unfolding preprocess. As an application, the GND-ICA can be used for multiple feature fusion and representation for color image classification. Many features extracted from a given image are constructed as a tensor. The effective components for each feature can be selected or extracted simultaneously and combined by using the GND-ICA.

Keywords: Feature Reduction and Manifold Learning, Machine Learning and Data Mining, Classification and Clustering
Abstract: A popular method for feature selection are filters based on the estimation of the mutual information between the features and the target. If the data is very high dimensional, even simple, iterative methods require a lot of computational time. In this work we propose an early stopping method for feature selectors that reduces the complexity of the feature selector by orders of magnitute without any loss of predictive performance. We demonstrate the practical use of early stopping on high dimensional image clasification tasks.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, Multimedia Analysis, Indexing and Retrieval
Abstract: With the explosive growth of Internet image data, labeling image data for image retrieval has become an increasingly onerous task. To that end, we proposed a novel multi-view learning with batch mode active learning framework, MV-BMAL, for improving the performance of image retrieval. Specifically, color, texture and shape features are extracted and considered as uncorrelated and sufficient views of an image, then each classifier is trained on these views respectively, and the schema makes full use of the classification results of each unlabeled samples to find out the most informative and representative samples for automatically or manually labeling. Finally, we evaluate MV-BMAL on benchmark data sets, and the experimental results show that our proposed MV-BMAL algorithm significantly outperforms the previous methods.

Keywords: Machine Learning and Data Mining
Abstract: Short message service (SMS) is now an indispensable way of social communication. However the mobile spam is getting increasingly serious, troubling users' daily life and ruining the service quality. We propose a novel approach for spam message detection based on mining the underlying social network of SMS activities. Comparing with strategies on keywords or flow detection, our network-based approach is more robust and difficult to defeat by human spammers. Various levels of features are employed to describe multiple aspects of the network, such as static structures, node activities and evolving situations. Experimental results on real dataset illustrate effectiveness of various features, showing our promising results.

Keywords: Pattern Recognition for Surveillance and Security, Pattern Recognition for Search, Retrieval and Visualization, Pattern Recognition for Bioinformatics
Abstract: Search optimization algorithms have the challenge of balancing between exploration of the search space (e.g., map locations, image pixels) and exploitation of learned information (e.g., prior knowledge, regions of high fitness). To address this challenge, we present a very basic framework which we call Zombie Survival Optimization (ZSO), a novel swarm intelligence approach modeled after the foraging behavior of zombies. Zombies (exploration agents) search in a space where the underlying fitness is modeled as a hypothetical airborne antidote which cures a zombie's aliments and turns them back into humans (who attempt to survive by exploiting the search space). Such an optimization algorithm is useful for search, such as searching an image for a pedestrian. Experiments on the CAVIAR dataset suggest improved efficiency over Particle Swarm Optimization (PSO) and Bacterial Foraging Optimization (BFO). A C++ implementation is available.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, Classification and Clustering
Abstract: We study the Chi-Squared distance and metric learning as a problem of Large Margin Nearest Neighbor (LMNN) classification. We suggest the Chi-Squared metric learning algorithm, based on the LMNN approach, to learn a metric to improve the accuracy of k-nearest neighbor (kNN) classification. We show that the Chi-Squared distance in the transformed space is one of the Quadratic-Chi distance family members. We use a gradient descent method to get an optimal value of the linear transformation matrix of the input feature space. We also show an alternative approach of the Chi-Squared metric learning by using a convex optimization method with relaxation. We demonstrate experimental results on the datasets mainly for domain adaptation.

Keywords: Pattern Recognition for Search, Retrieval and Visualization
Abstract: We propose an integrated and personalized video retrieval and summarization system. We estimate and impose appropriate preference values on affinity propagation graph of the video frames. Then, our system produces the summary which is useful for the user in her/his relevance feedback and for the retrieval module for comparing video pairs. The experiments confirm the effectiveness of our approach for various query types.

Keywords: Features and Image Descriptors
Abstract: We propose a novel local feature descriptor, Local Gaussian Directional Pattern (LGDP), for face recognition. LGDP encodes the directional information of the face's textures (i.e., the texture's structure) in a compact way, producing a more discriminating code than other methods. The structure of each micro-pattern is computed by using a derivative-Gaussian compass mask, and encoded by using its prominent directions and sign---which allows it to distinguish among similar structural patterns that have different intensity transitions. Moreover, our descriptor extracts several facial characteristics by varying the size of its mask, to recover features that may be missed in just one resolution. We construct the face descriptor by concatenating the LGDP's distributions extracted from a uniform grid of the face. We perform several experiments in which our descriptor performs consistently under illumination, noise, expression and age variations.

Keywords: 2D/3D Object Detection and Recognition, Biometrics, Pattern Recognition for Surveillance and Security
Abstract: One problem of existing appearance-based face recognition methods (e.g. PCA, LDA) is their weak ability of coping with local variations caused by facial expressions, motion deformation, data missing, etc. Multi-subregion fusion methods, which divide the face into a set of subregions, aim at this issue, and were reported a better performance. However, it leaves two open questions: 1. what subregions are good partitions on the face, and 2. How to fuse these subregions could achieve the expected performance. In this paper, we address these two questions and propose a local discrimination driven face partition method based on a coarse-to-fine Quad-tree decomposition. Unlike other multi-subregion approaches relying on prior knowledge, our method partitions the face according to the data property. Thus, it can adapt to varied databases. Meanwhile, our method introduces an optimized solution that fuses selected subregions to reach higher recognition accuracy. The cross-database experiments including one 3D database and three 2D databases demonstrate the efficiency and effectiveness of the proposed method.

Keywords: Motion, Tracking and Video Analysis, Stereo and Image-Based Modeling, Vision for Graphics
Abstract: This paper addresses the problem of estimating human gaze from eye appearance under free head motion. Allowing head motion remains challenging because eye appearance changes significantly for different head poses, and thus new head poses require new training images. To avoid repetitive training, we propose to produce synthetic training images for varying head poses. First, we model pixel displacements between head-moving eye images as 1D pixel flows, and then produce such flows to synthesize new training images from the original training images captured under a fixed default head pose. Specifically, we produce all the required 1D flows by using only four additionally captured images. Our method was successfully tested with extensive experiments to demonstrate its effectiveness.

Keywords: Motion, Tracking and Video Analysis
Abstract: We propose an approach for multi-pose face tracking by association of face detection responses in two stages using multiple cues. The low-level stage uses a two-threshold strategy to merge detection responses based on location, size and pose, resulting in short but reliable tracklets. The high-level stage uses different cues for computing a joint similarity measure between tracklets. The facial cue compares facial features of the most frontal face detections in pairs of tracklets. The classifier cue learns a discriminative appearance model for each tracklet, using detection pairs within reliable tracklets and between overlapping tracklets as training data. The constraint cue observes the compatibility of motion of two tracklets. The association of tracklets is globally optimized with the Hungarian algorithm. We validate our approach on two challenging episodes of two TV series and report a Multiple Object Tracking Accuracy (MOTA) of 82% and 68.2%, respectively.

Keywords: Gesture and Behavior Analysis, Motion, Tracking and Video Analysis
Abstract: Facial feature tracking and facial actions recognition from image sequence attracted great attention in computer vision field. Most current methods treat them as independent problems, hence ignore the interactions between facial feature points and facial actions. In this paper, we introduce a probabilistic framework based on the Dynamic Bayesian network (DBN) to simultaneously and coherently represent the facial evolvement in different levels, their interactions and their observations. Given the model and the measurements of facial motions, both facial features and facial actions are simultaneously recognized through probabilistic inference. Experiments show that compared to the state-of-the-art techniques, the proposed model can improve both the tracking and recognition performance.

Keywords: Stereo and Image-Based Modeling, Vision for Robotics
Abstract: The automatic surface reconstruction from an image sequence is still an active research topic. Recently, a method was designed to reconstruct a 2-manifold surface from the sparse cloud of points generated by Structure-from-Motion (SfM). This method reconstructs outdoor scenes from hundreds of catadioptric images. It is based on sculpting the 3d Delaunay triangulation of the points by ray-tracing. Our paper improves the resulting surface by removing spurious handles, which occur frequently.

Keywords: Stereo and Image-Based Modeling, Vision for Graphics
Abstract: This paper presents a novel parameterized variety based architecture for interactive 3DTV and Free viewpoint TV (FTV) applications. The proposed signal representation scheme allows to render free viewpoint images, taking only few sample images of the scene acquired by arbitrary, uncalibrated cameras. We extend the Parameterized Image Variety approach for rendering proposed earlier by Genc and Ponce [1] to handle full perspective cameras. A fast and efficient algebraic framework is proposed for the parameterized representation of 3D scene, in terms of image pixel positions corresponding to only three reference scene points. The key aspects of the novel FTV architecture based on this variety model are i) interactive stereoscopic view synthesis from arbitrary viewpoint ii) intuitive interface for content based virtual view specification iii) facilitation to add special effects like 3D scene augmentation.

Keywords: Stereo and Image-Based Modeling
Abstract: We propose a novel metric to measure the image-based 3D reconstruction results without ground truth datasets. In contrast to previous metrics, our shading-based metric can not only accurately measure the reconstruction quality but also decouple from any reconstruction algorithm. Considering the uncertainty of topology, texture and soft shadow, we compute an anisotropic irradiance gradient field from multiple images to indicate the regions where reconstruction error occurs. We further apply the metric into the view planning application. Experimental results on both synthetic and real datasets illustrate the effectiveness of evaluating 3D reconstruction by our metric. The reconstruction accuracy and completeness overtop or are the same as the results of manually adding new viewpoints.

Keywords: Scene Understanding, Stereo and Image-Based Modeling, Geometric and Photometric Registration
Abstract: Holistic scene understanding is a major goal in recent research of computer vision. To deal with this task, reasoning the 3D relationship of components in a scene is identified as one of the key problems. We study this problem in terms of structural reconstruction of 3D scene from single view image. Our first step concentrates on geometrical layout analysis of scene using low-level features. We allocate images into seven recurring and stable geometry classes. This classification labels the image with rough knowledge of its scene geometry. Then, based on this geometry label, we propose an adaptive autonomous scene reconstruction algorithm which adopts specific approaches particularly for different scene types. We show, experimentally, given the right geometry label, low-quality uncalibrated monocular images from the benchmark dataset can be structurally reconstructed in 3D space in a time/effort efficient way. This robust approach does not require a high quality or complex input image. We demonstrate the effectiveness of this approach in this paper.

Keywords: Stereo and Image-Based Modeling, Low-Level Vision, Classification and Clustering
Abstract: This paper presents a novel algorithm for estimating stereo disparity which exploits the benefit of learning to the fullest. Given a cost volume of stereo matching, we solve the cost aggregation and disparity computation in on shot by using a classifier; we design a feature called matching cost pattern, for the input which we extract from the cost volume while we use simulated stereo patterns for training. To this end, we introduce a highly realistic computer graphics dataset, the new Tsukuba stereo dataset, with ground-truth disparity maps. Through preliminary experiments we show that our algorithm outperforms a simplified AD-Census cost-minimization method, and also that the error ratio decreases as we use a larger number of samples for training.

Keywords: Image and Video Processing, Enhancement, Restoration and Filtering
Abstract: In this paper, a novel image fusion based interpolation method is proposed. Soft-decision-adaptive interpolation (SAI) is one of the state of the art image interpolation algorithms. However, SAI may produce serious artifacts in small -scale edge areas. Bicubic interpolation performs better in preserving the fidelity of small -scale edges. But, bicubic interpolat ion may cause zigzag and blurring artifacts around strong edges. The proposed method combines the advantages of SAI and Bicubic together thro ugh image fusion. The artifacts in the SAI interpolated image are first detected and then removed by fusing the SAI interpolated image with the bicubic interpolated image. Experiments demonstrate the effectiveness of the proposed method.

Keywords: Image and Video Processing, Enhancement, Restoration and Filtering
Abstract: This paper presents a variational framework for obtaining super-resolved video-sequences, based on the observation that reconstruction-based Super- Resolution (SR) algorithms are limited by two factors: registration exactitude and Point Spread Function (PSF) estimation accuracy. To minimize the impact of the first limiting factor, a small-scale linear inpainting algorithm is proposed to provide smooth SR video frames. To improve the second limiting factor, a fast PSF local estimation and total variation-based denoising is proposed. Experimental results reflect the improvements provided by the proposed method when compared to classic SR approaches.

Keywords: Image and Video Processing, Enhancement, Restoration and Filtering
Abstract: This paper presents a new approach to image superresolution based on sparse representation. This problem is formulated as a compressive sensing system, in which an over-complete dictionary is used to sparsely represent a low resolution image and generate a high resolution image. We propose a method to adaptively construct training images by selecting the most correlative images based on scale-invariant feature transform(SIFT). In this way, the sampling and dictionary training process is accelerated and optimized. In comparison with the existing approaches, experimental results show that the proposed method outperforms the existing benchmark in terms of both super-resolution quality and the processing time, which makes the propose method suitable for practical applications.

Keywords: Enhancement, Restoration and Filtering, Features and Image Descriptors, Image and Video Processing
Abstract: Edges are of significant importance in visual resolution perception. In this paper, we propose a novel image super-resolution method by enhancing the edges in the low resolution image. We first define a new edge sharpness feature: gradient profile sharpness (GPS), which considers both the absolute magnitude and the spatial scattering of edge gradient profile. Then we learn the relationship between GPSs in high resolution images and low resolution images, and we formulate a linear GPS transform to provide gradient prior for image reconstruction. Our GPS can represent edge sharpness perceptually well. And our super-resolution method can output harmonious and faithful images with better reconstruction quality.

Keywords: Enhancement, Restoration and Filtering, Image and Video Processing
Abstract: Super-resolution using web images is a challenging problem because accurate motion estimation is difficult due to their viewpoint change and photometric variation. To overcome such difficulties, this paper formulates a unified framework where local color correction is incorporated into fuzzy motion estimation after global registration. Thus, both geometric and photometric variations are simultaneously taken into account for implicit motion estimation. As demonstrated in the results using web images under various conditions, the proposed approach magnifies user-selected region of interest with visually appealing details.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Text classification (TC) has long been an important research topic in information retrieval (IR) related areas. Conventional language model (LM)-based TC is solely based on matching the words in the documents and classes by using a naive Bayes classifier (NBC). In the literature, both the term association model (TA), which further considers word-to-word information, and the relevance model (RM), which further considers word-to-document information, have been shown to outperform a simple LM for IR. In this paper, we study a novel integration of TA with RM for LM-NBC-based TC. The new model is called the term relevance dependency model (TRDM). In the model, the probability of a word given a class is represented by a term association LM probability learned by a RM framework. The results of TC experiments on the 20newsgroups and Reuters-21578 corpora demonstrate that the new model outperforms the standard NBC and several other LM-NBCbased methods.

Keywords: Classification and Clustering, Neural Networks, Machine Learning and Data Mining
Abstract: Diversity is a key element in the success of classifier ensembles, and has attracted much recent attention. It is typically measured by directly computing the amount of disagreement between ensemble classifiers at the decision level. This costly process usually involves evaluating output predictions of each classifier over some validation data set. Since most statistical and neural network classifiers can adjust internal learning dynamics by varying their hyperparameter values (corresponding to genotype values), this information can also provide an estimate of diversity. This paper measures the correlation between genotype and classifier diversity among an ensemble of fuzzy ARTMAP neural network classifiers applied to video face recognition. It is empirically shown that as genotype diversity increases within an ensemble, classifier diversity also significantly increases. This correlation can then be exploited to measure the diversity among base classifiers during ensemble design with a significantly lower computational cost.

Keywords: Classification and Clustering, Machine Learning and Data Mining
Abstract: Motivated by the observation that communities in real world social networks form due to actions of rational individuals in networks, we propose a novel game theory inspired algorithm to determine communities in networks. The algorithm is decentralized and only uses local information at each node. We show the efficacy of the proposed algorithm through extensive experimentation on several real world social network data sets.

Keywords: Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining
Abstract: In this paper a new variation of Support Vector Machines (SVM) is introduced. The proposed method is called Subclass Support Vector Machine (SSVM) and makes use of principles from Discriminant Analysis field using subclasses. The major difference over SVM is that it takes into account the existence of subclasses in the classes and tries to minimize the distribution of the samples within each subclass. Experiments over various databases are conducted and the results are compared against other classifiers.

Keywords: Statistical, Syntactic and Structural Pattern Recognition
Abstract: We present a new approach to compute the graph edit distance between two attributed graphs which is based on a formal connection between the graph edit distance problem and that of finding a dominant set in an auxiliary edge-weighted "association" graph. Experiments performed on various data sets show that with the proposed approach we are able to improve on state-of-the-art algorithms.

Keywords: Gesture and Behavior Analysis, Classification and Clustering, Machine Learning and Data Mining
Abstract: Classifying continuous signals from multiple channels poses several challenges: different sample rates from different types of channels have to be incorporated. Furthermore, when leaping from the laboratory to the real world, it is mandatory to deal with failing sensors and also uncertain or even incorrect classifiations. We propose a new Multi Classifier System (MCS) based on the application of classifier making use of an reject option and a Markov Fusion Network (MFN) which is evaluated in an off-line and on-line manner. The architecture is tested using the publicly available AVEC corpus, that collects affectively labeled episodes of human computer interaction. The MCS achieved a significant improvement compared to the results obtained on the single modalities.

Keywords: Pattern Recognition for Surveillance and Security
Abstract: There are cases where surveillance video must be transmitted in the clear, which is without encryption, such as when multiple public safety agencies use the same video feeds. To ensure trust of the received video, it must be authenticated. While lossless video is straightforward to authenticate by cryptographic means, lossy video as may result from UDP, wireless, or transcoded transmissions, is more difficult to authenticate. We describe a method that combines a concise and efficiently computed video fingerprint with public key cryptography to enable lossy video authentication. Experimental results show how video fingerprint length relates to successful authentication rate and latency of its calculation.

Keywords: Pattern Recognition for Surveillance and Security, Image and Video Understanding, Human Computer Interaction
Abstract: We propose a novel human-in-the-loop surveillance system that continuously learns the properties of objects that are interesting for a human operator. The interesting objects are automatically learned by tracking the eye gaze positions of the operator while he or she monitors the surveillance video. The system automatically detects interesting objects in the surveillance video and forms a new synthetic video that contains interesting objects at earlier positions in the time dimension. The operator always views this synthetically formed video which makes manual video retrieval tasks more convenient. Sensitivity to operator interests and interest changes are other major advantages. We tested our system both on synthetic and real videos, which are provided as supplementary materials. The results show the effectiveness of the proposed system.

Keywords: Pattern Recognition for Art, Cultural Heritage and Entertainment, Multimedia Analysis, Indexing and Retrieval, Pattern Recognition for Search, Retrieval and Visualization
Abstract: In this paper, we suggest an interaction model designed to fit users' expectations in front of an image retrieval system. A lightweight relevance feedback strategy, working directly on the 2D projection of image features, allows the user to spatially navigate the media collection maintaining the real-time constraint. A preliminary evaluation of this relevance feedback strategy shows good performance compared with other known approaches.

Keywords: Pattern Recognition for Surveillance and Security, Speech and Audio Processing, Image and Video Processing
Abstract: We present a state-of-the-art bi-modal authentication system for mobile environments, using session variability modelling. We examine inter-session variability modelling (ISV) and joint factor analysis (JFA) for both face and speaker authentication and evaluate our system on the largest bi-modal mobile authentication database available, the MOBIO database, with over 61 hours of audio-visual data captured by 150 people in uncontrolled environments on a mobile phone. Our system achieves 2.6% and 9.7% half total error rate for male and female trials respectively � relative improvements of 78% and 27% compared to previous results.

Keywords: Gesture and Behavior Analysis, Image and Video Understanding
Abstract: This paper proposes a new method for facial motion extraction to represent, learn and recognize observed expressions, from 4D video sequences. The approach called Deformation Vector Field (DVF) is based on Riemannian facial shape analysis and captures densely dynamic information from the entire face. The resulting temporal vector field is used to build the feature vector for expression recognition from 3D dynamic faces. By applying LDA-based feature space transformation for dimensionality reduction which is followed by a Multiclass Random Forest learning algorithm, the proposed approach achieved 93% average recognition rate on BU-4DFE database and outperforms state-of-art approaches.

Keywords: Gesture and Behavior Analysis
Abstract: In this paper, we try to distinguish between the posed and spontaneous facial expressions through infrared thermal images. Firstly, six statistical features are ex- tracted from the temperature difference matrices be- tween the apex and onset facial expression frames in each facial subregion. Secondly, the thermal differ- ences between the posed facial expressions and spon- taneous ones are analyzed by a Paired-Samples T- test. Thirdly, Bayesian Networks are adopted to clas- sify posed and spontaneous expressions based on these temperature statistical features. Experimental results on the thermal subdatabase of the USTC-NIVE demon- strate the feasibility and effectiveness of posed and spontaneous expression recognition through facial tem- perature changes. This study also illustrates that the importance of different facial subregions� temperature variances for differentiating different kinds of posed and spontaneous expressions.

Keywords: Image and Video Understanding, Neural Networks, Motion, Tracking and Video Analysis
Abstract: Face alignment is a crucial step in many facial processing applications. It has received extensive attention in the last two decades. The general approach consist in estimating the parameters of a deformable shape model, which minimize a cost function. Most of the existing methods are based on empirical cost functions. In this paper we propose to learn an ideal global cost function (i.e. the quality of the alignment) as convex as possible in order to lead to a simple and robust optimization step. This learning process relies on the Boosted Input Selection Algorithm for Regression (BISAR). It selects the best set of Haar-like features as input of a Neural Network to predict the value of the cost function. Performance of this method is evaluated on unseen data from the training database. The generalization performance is assessed on unseen data from unrelated datasets. This approach is also favorably compared with a state-of-the-art method.

Keywords: Image and Video Processing, Detection, Separation and Segmentation, Geometric and Photometric Registration
Abstract: We extend the PCT (Pseudo Census Transform)-based appearance model to ranking-based appearance model for face alignment. The PCT-based weak ranking function is learned using RankSVM, and the ranking appearance model (RAM) is constructed in a boosting manner. Experiments show that the PCT-based RAM is more robust and generalize better than the PCT-based boosted appearance model (BAM). The PCT-RAM achieves about 23% improvement when tested on unseen data. We also investigate different sampling strategies for the learning to rank problem and find out that random permutation achieves similar results as using adjacent ordering pairs. The alignment results do not decrease significantly when only one ordinal pair is used for each direction.

Keywords: Image and Video Understanding, Speech and Audio Processing, Classification and Clustering
Abstract: We present a novel facial expression recognition framework using audio-visual information analysis. In particular, we design a single good image representation of the image sequence by weighted sum of registered face images where the weights are derived using auditory features. We use a still image based technique for the expression recognition task. We performed experiments using eNTERFACE'05 audio-visual emotional database. The analysis shows that our framework can improve the recognition performance while significantly reducing the computational cost by avoiding redundant or insignificant frame processing by incorporating auditory information.

Keywords: Gesture and Behavior Analysis
Abstract: Not only facial expressions but also body gestures and postures play an important role in non-verbal communication. Facial expressions are based on two factors: arousal and pleasant emotions, while it is not clear that body gestures and postures have the same structure as the facial expressions have. We indicate that (1) the sitting postures have the same emotion structure as facial expressions and (2) can be measured by pressure sensors on a chair and accelerometers on the body, which predict the emotion factors. We find the sitting postures have the semantic factors: "arousal", "pleasantness", and "avoidance", so emotion expressions of the sitting postures are similar to those of the facial expressions. Their difference is "avoidance" expressed by not the main body but the body parts such as arms and legs. We conclude that (1) "arousal" and "pleasantness" factors can be measured with the proposed sensors and (2) the body trunk and the body parts: neck, arms, and legs are important.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Classification and Clustering
Abstract: Mixture models are frequently used to classify data. They are likelihood based models, and the maximum likelihood estimates of parameters are often obtained using the expectation maximization (EM) algorithm. However, multimodality of the likelihood surface means that poorly chosen starting points for optimisation may lead to only a local maximum, not a global maximum. In this paper, different methods of choosing starting points will be evaluated and compared, mainly via simulated data and then we will introduce a procedure which will make intelligent choices of possible starting points and fast evaluations of their usefulness.

Keywords: Classification and Clustering, Camera-Based Document Analysis, Image and Video Understanding
Abstract: A recognition method of road markings for map generation is presented. For accurate position estimation and classification, two voting schemes are proposed and　combined. The first is multi-frame sparse tensor voting　for geometric feature extraction, and the second is contour　localization using the resulting tensor field. Classification is based on the similarity between the　aligned contour and the tensor field. The experimental results show that the proposed method outperforms conventional matching-based approaches.

Keywords: Biometrics, Feature Reduction and Manifold Learning
Abstract: Feature selection is an important issue in pattern recognition. In face recognition, one of the state-of-the-art methods is that some feature selection methods (e.g., AdaBoost) are first utilized to select the most discriminative features and then the subspace learning methods (e.g., LDA) are further applied to learn the discriminant subspace for classification. However, in these methods, the objective of feature selection and subspace learning is not so consistent and the combination is not the optimal. In this paper, we propose a novel and efficient feature selection method that is designed for linear discriminant analysis (LDA). We use the Fisher criterion to select the most discriminative and appropriate features so that the objectives of feature selection and classifier learning are consistent (both follow the Fisher criterion) and the face recognition performance is expected to be improved. Experiments on FRGC v2.0 face database validate the efficacy of the proposed method.

Keywords: Machine Learning and Data Mining, Classification and Clustering
Abstract: While there is a lot of research on change detection based on the streaming classification error, finding changes in multidimensional unlabelled streaming data is still a challenge. Here we propose to apply principal component analysis (PCA) to the training data, and mine the stream of selected principal components for change in the distribution. A recently proposed semi-parametric log-likelihood change detector (SPLL) is applied to the raw and the PCA streams in an experiment involving 26 data sets and an artificially induced change. The results show that feature extraction prior to the change detection is beneficial across different data set types, and specifically for data with multiple balanced classes.

Keywords: Gesture and Behavior Analysis, Pattern Recognition for Surveillance and Security
Abstract: Many accidents at intersection are happend due to judgment errors, which are caused from blind spot of drivers. We propose a method for generating warning messages related to the estimated accidents between blind spots of drivers and vehicle behaviors by using surveillance camera and on-board camera frames. In this method, vehicle behaviors of straight, right turn, left turn, change to right lane, and change to left lane are detected by applying hidden markov model (HMM). Direction sequences of behaviors are used for training HMMs which represent each behavior by Baum-Welch algorithm. Behavior information enables the drivers to recieve our intelligent support and allows drivers to recognize behaviors of other vehicles easily. From a result of behavior detection, our system could recognize 80% of behavior correctly.

Keywords: Gesture and Behavior Analysis, Classification and Clustering, Human Computer Interaction
Abstract: This paper presents a novel approach for automatic audio-visual emotion recognition. The audio and visual channels provide complementary information for human emotional states recognition, and we utilize coupled HMM (CHMM) as model-level fusion method in our work. To further improve recognition accuracy, we design an AdaBoost-CHMM ensemble classifier which takes CHMM as component classifiers in adaptive boosting procedure. A modified expectation-maximization (EM) algorithm for CHMM learning is proposed to make the learning process focus more on difficult samples. Experiment results on audio-visual emotion data collected in Wizard of Oz scenarios and labeled under two types of emotion category sets demonstrate that our approach is effective and promising.

Keywords: Classification and Clustering, Occlusion and Shadow Detection, Scene Understanding
Abstract: Most of conventional object matching methods are based on comparing the local features, which are too computational demanding to achieve realtime performance on object detection in videos. Recently, Dominant Orientation Templates (DOT) method was proposed to make online feature detection and comparison feasible. However, it still suffers the problem of fragility due to the noise and partial occlusions. To efficiently tackle these problems, we introduce the similarity map to store the matching scores of individual grids in each sliding window, which is used to further denoise and infer the occlusion map. The promising experimental results demonstrate that proposed approach improves the robustness, which outperforms DOT assuredly.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Machine Learning and Data Mining, Classification and Clustering
Abstract: Multiple Instance Learning (MIL) is concerned with learning from sets (bags) of objects (instances), where the individual instance labels are ambiguous. In MIL it is often assumed that positive bags contain at least one instance from a so-called concept in instance space. However, there are many MIL problems that do not fit this formulation well, and hence cause traditional MIL algorithms, which focus on the concept, to perform poorly. In this work we show such types of problems and the methods appropriate to deal with either situation. Furthermore, we show that an approach that learns directly from dissimilarities between bags can be adapted to deal with either problem.

Keywords: Gesture and Behavior Analysis, Motion, Tracking and Video Analysis
Abstract: In traditional bag-of-words method, each local feature is treated evenly for representation. One disadvantage of this kind of strategy is that it is not robust to noise, which makes the performance impaired. In this paper, a novel human action recognition approach which learns weights for features is proposed, where each feature is assigned a weight for human action representation. In our proposed strategy, the weights and the discriminative model are learned jointly. There are two advantages of our model. First, small weights are assigned to noise, which can help to reduce the effect of noise in the representation process. Second, discriminative features, which are critical for human action recognition, are assigned large weights. Experimental results demonstrate the advantages of our method.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Features and Image Descriptors, Pattern Recognition for Search, Retrieval and Visualization
Abstract: In this paper we tackle the problem of subgraph isomorphism detection on large graphs, which may commonly be intractable, even with state of the art algorithms. Rather than competing with other matching algorithms, we define enhancements that can be used by (almost) any subgraph isomorphism algorithm, both current and future. These enhancements consist of a number of topological features to be added to the nodes, and a technique which we term "iterative node elimination". The fusion of these enhancements is shown to be able to reduce subgraph isomorphism matching times by a factor of over 4,500.

Keywords: Machine Learning and Data Mining
Abstract: This paper describes an textbf{active transfer learning} technique for multi-view head-pose classification. We combine transfer learning with active learning, where an active learner asks the domain expert to label the few textbf{most informative} target samples for transfer learning. Employing adaptive multiple-kernel learning (AMKL) for head-pose classification from four low-resolution views, we show how active sampling enables more efficient learning with few examples. Experimental results confirm that active transfer learning produces 10% higher pose-classification accuracy over several competing transfer learning approaches.

Keywords: Pattern Recognition for Surveillance and Security, Motion, Tracking and Video Analysis, Scene Understanding
Abstract: This paper addresses the problem of identification of pedestrian groups in crowded environments. To that end, positional and directional relations are modeled accounting for different environmental features and group configurations. Subsequently, a pair of simultaneously observed pedestrians is identified to belong to the same group or not utilizing these models in a parallel manner, which defines a compound hypothesis testing scheme. In case of ambiguities, local and global indicators of group relation are employed in quantifying the reliabilities of the two individual decisions. The contribution of this study lies in the improvement in positional and directional relation models to adjust to different environments and group configurations, description of compound evaluation, as well as resolution of ambiguities proposing uncertainty measures based on local and global indicators of group relation.

Keywords: Pattern Recognition for Surveillance and Security, Motion, Tracking and Video Analysis, Segmentation, Color and Texture
Abstract: A novel two-stage background generation method is proposed. In the first stage, an intensity-level based statistical approach is employed to identify a variety of background variations. No background training is needed. In the second stage, a background variation based heuristic framework is designed to generate a synchronized background video sequence on-line from the surveillance video. This framework employs both static and dynamic spatial clues in the scene. The quantitative evaluations demonstrate the method can generate more realistic background images than some well-known background modeling techniques.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Pattern Recognition for Bioinformatics
Abstract: In the domain adaptation research, which recently becomes one of the most important research directions in machine learning, source and target domains are with different underlying distributions. In this paper, we propose an ensemble learning framework for domain adaptation. Owing to the distribution differences between source and target domains, the weights in the final model are sensitive to target examples. As a result, our method aims to dynamically assign weights to different test examples by making use of additional classifiers called model-friendly classifiers. The model-friendly classifiers can judge which base models predict well on a specific test example. Finally, the model can give the most favorable weights to different examples. In the experiments, we firstly testify the need of dynamical weights in the ensemble learning based domain adaptation, then compare our method with other classical methods on real datasets. The experimental results show that our method can learn a final model performing well in the target domain.

Keywords: Feature Reduction and Manifold Learning, Statistical, Syntactic and Structural Pattern Recognition, Neural Networks
Abstract: An appropriate dimension reduction of raw data helps to reduce computational time and to reveal the intrinsic structure of complex data. In this paper, a dimension reduction method for regression is proposed. The method is based on the well-known sliced inverse regression and conditional entropy minimization. Using entropy as a measure of dispersion of data distribution, dimension reduction subspace is estimated without assuming regression function form nor data distribution, unlike conventional sliced inverse regression. The proposed method is shown to perform well compared to some conventional methods through experiments using both artificial and real-world data sets.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition, Classification and Clustering
Abstract: Kernel methods have been successfully used in many practical machine learning problems. However, the problem of choosing a suitable kernel is left to practitioners. One method to select the optimal kernel is to learn a linear combination of element kernels. A framework of multiple kernel learning based on conditional entropy minimization criterion (MCEM) has been proposed and it has been shown to work well for, e.g., speaker recognition tasks. In this paper, a computationally efficient implementation for MCEM, which utilizes sequential quadratic programming, is formulated. Through a comparative experiment to conventional MCEM algorithm on a speaker verification task, the proposed method is shown to offer comparable verification accuracy with considerable improvement in computational speed.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Pattern Recognition for Search, Retrieval and Visualization
Abstract: In this paper, we propose a novel method called textbf{THUNTER} to textbf{t}ransfer the textbf{h}eterogenous textbf{un}labeled data in the source domain to the target domain for clustextbf{ter}ing. Suppose the target data are a set of images, then the so-called heterogeneous unlabeled data can be a large set of text data or acoustic data. Our method aims to address how to transfer these large amount of heterogeneous data to the relatively smaller set of target data for clustering. To the best of our knowledge, it is the first work in the community to transfer the unlabeled data, especially the unlabeled heterogeneous data, for unsupervised clustering. Furthermore, along with our method, a novel textbf{dic}tionary-based data textbf{trans}fer strategy (textbf{DicTrans}) is introduced in this paper, which measures the fidelity of transferring the data in the source domain and automatically decides how many to transfer. Through a series of experiments, the effectiveness of THUNTER and DicTrans are demonstrated with very promising performances.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, Image and Video Processing, Character and Text Recognition
Abstract: In this paper, we present an effective approach to locate scene text in images based on connected components analysis (CCA). Our approach first utilizes a multi-scale adaptive local thresholding operator to convert an image into two complementary binary images. Then, connected components (CCs) are extracted from both of them, which ensures that bright or dark text in contrast to background can be detected. Further, some rules are designed based on stroke features to verify whether a connected component belongs to characters, and the obtained candidate components are further checked on the word level by using a graph to represent spatial relation of different components. Finally, scene text regions are localized by searching the collinear maximum group over the graph. The comparison experiments of the proposed method with some representative state-of-the-art methods, on the challenging subset of ICDAR 2003, show that the proposed approach is very effective, and it is robust to text of different sizes, fonts,colors, as well as orientation of text lines.

Keywords: Biometrics, Pattern Recognition for Surveillance and Security, Image and Video Processing
Abstract: While identical twins identification is a well known challenge in face recognition, it seems that no work has explored automatic ear recognition for identical twin identification. Ear image recognition has been studied for years, but Iannarelli (1989) appears to be the only work mentioning the twin identification (performed manually). We here explore the possibility of automatic twin identification from their ear images based on a recently proposed psychological model for face recognition in humans, known as Exception Report Model (ERM). We test our new approach on 39 pairs of identical twins (78 subjects), with several levels of resolution, occlusion and noise, left ear vs. right ear, and feature optimization which verifies the robustness of the introduced features.

Keywords: Feature Reduction and Manifold Learning, Classification and Clustering, Machine Learning and Data Mining
Abstract: Traditional nonlinear feature selection methods map the data from an original space into a kernel space to make the data be separated more easily, then move back to the original feature space to select features. However, the performance of clustering or classification is better in the kernel space, so we are able to select the features directly in the kernel space and get the direct importance of each feature. Motivated by this idea, we propose a novel method for unsupervised feature selection directly in the kernel space. To do this, we utilize local discriminative information to find the best label for each instance with L2,1-norm minimization, then select the most important features in the kernel space using the labels predicted. Extensive experiments demonstrate the effectiveness of our method.

Keywords: Biometrics, Motion, Tracking and Video Analysis, Gesture and Behavior Analysis
Abstract: We describe a new method to relieve common assumptions/ restrictions in head tracking by using a model-based approach. This improves local feature matching which only considers the pattern around the extracted feature excluding the object shape, so that misalignment can occur. In this paper, to overcome constraints on motion we consider region- and distance-based feature refinement methods to validate the local features used when tracking the ellipsoidal object. We also present a direct mapping method to reconstruct 3D feature positions for tracking. The utility of the new method has been demonstrated for face pose estimation using the Boston face database.

Keywords: Feature Reduction and Manifold Learning, Biometrics, Machine Learning and Data Mining
Abstract: The notion of metric is fundamental for the study of pattern recognition and vector 2-norm |cdot|_{2} is one of the most widely used metric, i.e., Euclidean distance. However, there is often the case that the inputs are matrices, e.g., 2D images in face recognition. Since a matrix can take more structure information than its vectorization, it is highly preferable to adopt the matrix representation of the original image rather than a simple vector.

In this paper, we first propose a class of discriminative metrics for matrices, i.e., Schatten p-norm, by which we can better explain that with Euclidean metric, why the differences among facial images due to impact factors, e.g., illuminations, are more significant than differences due to identity variations. Second, we propose a novel Principal Component Analysis method based on Schatten 1-norm which can be easily extended to other subspace learning methods. Extensive experiments on Yale B, CMU PIE, ORL and AR databases prove the effectiveness of our method.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Pattern Recognition for Search, Retrieval and Visualization
Abstract: This paper presents a novel algorithm for localizing and extracting the four clock-digits by using the knowledge of the digital video clock. This algorithm has triple merits compared with the existing ones. First, it is novel and easy implementation. It uses the pixel second-periodicity method. By using this method the algorithm omits the several tedious image processing steps. Second, it is faster than the existing algorithms. Third, it has good performance in both localization and extraction. The good performance is due to that it uses the domain knowledge in acquiring the digits color and in modeling the clock-digits layout. Experimental results show that our algorithm achieves an accuracy of 100% in localizing the four clock-digits at a very low computing cost and extracts the four clock-digits in a high accuracy.

Keywords: Classification and Clustering, Image and Video Understanding
Abstract: Insect species recognition is more difficult than generic object recognition because of the similarity between different species. In this paper, we propose a hybrid approach called discriminative local soft coding (DLSoft) which combines local and discriminative coding strategies together. Our method takes use of neighbor codewords to get a local soft coding and class specific codebooks (sets of codewords) for a discriminative representation. On obtaining the vector representation of image via spatial pyramid pooling of patches, a linear SVM classifier is used to classify images into species. Experimental results show that the proposed method performs well on insect species recognition and outperforms the state-of-the-art methods on generic object categorization.

Keywords: Biometrics, Statistical, Syntactic and Structural Pattern Recognition, Detection, Separation and Segmentation
Abstract: This paper presents new fingerprint minutiae detection by the advanced ridge flux analysis. The considerable processing time taken by the conventional approaches, most of which use the ridge thinning process with a rather large calculation time, is a problem that has recently attracted increased attention. Though Ridge flux analysis method without using thinning process is proposed in order to reduce the computational time, there still remains a problem with low detection accuracy. The proposed approach is applied to detect minutiae by analyzing the curvature of ridge contours to achieve both of the computation time reduction and higher detection accuracy. The experimental results show that the proposed approach can achieve a reduction in calculation time, while achieving the same success detection rate as that of the conventional approaches.

Keywords: Machine Learning and Data Mining, Classification and Clustering
Abstract: The large amount of digital data requests for scalable tools like efficient clustering algorithms. Many algorithms for large data sets request linear separability in an Euclidean space. Kernel approaches can capture the non-linear structure but do not scale well for large data sets. Alternatively, data are often represented implicitly by dissimilarities like for protein sequences, whose methods also often do not scale to large problems. We propose a single algorithm for both type of data, based on a batch approximation of relational soft competitive learning, termed fast generic soft-competitive learning. The algorithm has linear computational and memory requirements and performs favorable to traditional techniques

Keywords: Classification and Clustering, Machine Learning and Data Mining
Abstract: In this paper we present an optimization of the Optimum-Path Forest classifier training procedure, which is based on a theoretical relationship between minimum spanning forest and optimum-path forest for a specific path-cost function. Experiments on public datasets have shown that the proposed approach can obtain similar accuracies to the traditional one, but faster for data training.

Keywords: Classification and Clustering, Machine Learning and Data Mining, Feature Reduction and Manifold Learning
Abstract: Recent years have witnessed a growing interest in the sparse representation problem. Prior work demonstrated that adaptive dictionary learning techniques can greatly improve the performance of sparse representation approaches. Existing techniques mainly focus on the reconstructive accuracies and the discriminative power of the learned dictionary, whereas the mutual incoherence between any two basis atoms has been rarely studied yet. This paper proposes a novel method by explicitly incorporating a correlation penalty into the dictionary learning model. Experiments show that the proposed method can remarkably reduce the correlation measure of the learned dictionaries, and at the same time achieve higher classification accuracies than state-of-the-art algorithms.

Keywords: Classification and Clustering, Pattern Recognition for Bioinformatics
Abstract: In this paper, we propose a novel sparse representation method for classification called similarity weighted sparse representation (SWSR). The similarity weighted ℓ1-norm minimization, where the weighted matrix is constructed by incorporating the similarity information between the test sample and the entire training samples, is presented as an alternative to ℓ0-norm minimization to seek the optimal sparse representation for the test sample in SWSR. The sparse solution of SWSR encodes more discriminative information than other competing alternatives to ℓ0- norm minimization, so it is more suitable for classification. The experimental results on publicly available face databases demonstrate the efficacy of the proposed method.

Keywords: Pattern Recognition for Surveillance and Security, Motion, Tracking and Video Analysis, Classification and Clustering
Abstract: This paper proposes a novel nonlinear manifold learning method for addressing the ill-posed problem of occluded human action analysis. As we know, a person can perform a broad variety of movements. To capture the multiplicity of a human action, this paper creates a low-dimensional manifold for capturing the intra-path and inter-path contexts of an event. Then, an action path matching scheme can be applied for seeking the best event path for linking the missed information between occluded persons. After that, a recovering scheme is proposed for repairing an occluded object to a complete one. Then, each action can be converted to a series of action primitives through posture analysis. Since occluded objects are handled, there will be many posture-symbol- converting errors. Instead of using a specific symbol, we code a posture using not only its best matched key posture but also its similarities among other key postures. Then, recognition of an action taken from occlude objects can be modeled as a matrix matching problem. With the matrix representation, different actions can be more robustly and effectively matched by comparing their Kullback�Leibler(KL) distances.

Keywords: Pattern Recognition for Bioinformatics, Classification and Clustering, Biometrics
Abstract: It is widely accepted that biometric authentication systems based on human faces are vulnerable to spoofing attacks, in which attackers exploit recaptured photos of legitimate users to gain unauthorized access. To address this vulnerability, numerous approaches which employ complex physics-based models to verify the liveness of presented images have been proposed. While capable of distinguishing between real and recaptured images, a key weakness of these methods is their reliance on an appropriate choice of the model used. Here, we adopt a fundamentally different approach to detect recaptured images. Our method uses randomized temporal affective cues in the form of facial expressions to verify the liveness of users. A quantitative evaluation user study involving 6 users demonstrated the feasibility of our method, and we have also shown that the system is capable of achieving an average classification accuracy of 95.85% for different facial expressions.

Keywords: Feature Reduction and Manifold Learning, Multimedia Analysis, Indexing and Retrieval
Abstract: KNN-based image annotation method is proved to be very successful. However, it suffers from two issues: (1) high computational cost; (2) the difficulty of finding semantically similar images. In this paper, we propose a graph-based dimensionality reduction method to solve the two problems by adapting the locality sensitive discriminant analysis method [1] to multi-label setting. We first determine relevant and irrelevant images based on label information and construct relevant and irrelevant graphs by focusing on the visually similar relevant and irrelevant images. A linear feature transformation matrix is derived by considering the two graphs. The transformation can map the images to a low-dimensional subspace in which neighborhood relevant images are pulled closer while irrelevant images are pushed away. Thus the new feature after dimensionality reduction is quite fit for KNN-based image annotation. Experiments on the Corel dataset also demonstrate the effectiveness of our dimensionality reduction method for KNN-based image annotation.

Keywords: Biometrics, Features and Image Descriptors, Classification and Clustering
Abstract: Automatic age classification from human faces is a challenging task which has recently attained an increasing attention. Most of the proposed approaches have however been mainly concerning controlled settings. In this paper, we propose a novel method for age classification in unconstrained conditions and provide extensive performance evaluation on benchmark datasets with standard protocols, thus allowing a fair comparison and an easy reproduction of the results. Our proposed method is based on a combination of local binary pattern (LBP) variants encoding the structure of elongated facial micro-patterns and their strength. The experimental analysis points out the complexity of the age classification problem under uncontrolled settings. The proposed method provides state-of-the-art performance that can be used as a reference for future investigations.

Keywords: Machine Learning and Data Mining, Statistical, Syntactic and Structural Pattern Recognition, Classification and Clustering
Abstract: This paper explores the classification problem based on parallel feature partitioning. This formulation leads to a new problem in computational geometry. While this new problem appears to be NP-complete, it is shown that the proposed graph theoretical platform makes it semi-tractable, allowing the use of conventional tools for its solution. Here, by conventional, we mean any exact or heuristic algorithm for partitioning a graph into a minimal number of cliques or for finding the clique of maximum cardinality while seeking an efficient heuristic algorithm. An important advantage of this approach is the decomposition of a problem involving l classes into l optimization problems involving a single class. The computational complexity of the method, computational procedures, and classification rules are discussed. A geometrical interpretation of the solution is also given. Using the proposed approach, the geometrical structure of the training set is utilized in the best possible way.

Keywords: Feature Reduction and Manifold Learning, Classification and Clustering, Machine Learning and Data Mining
Abstract: This paper presents an approach for computing global distance metrics that minimize the k-NN leave-one-out (LOO) error. The approach optimizes an energy function that corresponds to a smoothened version of the k-NN LOO error. The generalization of the proposed approach is further improved by controlling the k parameter through a heuristic. Evaluation of the proposed approach on several public datasets showed that it was able to compete with an established state-of-the art approach.

Keywords: Pattern Recognition for Bioinformatics, Pattern Recognition for Surveillance and Security, Motion, Tracking and Video Analysis
Abstract: In this paper, we present a new method to generate a 3D face model, based on both Data-Driven and Structure-from-Motion approach. Considering both 2D frontal face image constraint, 3D geometric constraint, and likelihood constraint, we are able to reconstruct subject�s face model accurately, robustly, and automatically. Using our method, it is possible to create a 3D face model in 5.8 [sec] by only shaking own head freely in front of a single video camera.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, Detection, Separation and Segmentation, Image and Video Processing
Abstract: We have been manufacturing our own License Plate Recognition (LPR) systems since 1999 for parking lots in Japan. We have implemented a number of enhancements for our latest LPR system and one of the most important enhancements is for ranges of acceptable license-plate attitude angles. We collected 163,574 actual images so far and we selected 6,648 images randomly from this image database to verify our latest design. The histograms of plate attitude angles are estimated for these 2,836 images by measuring the license-plate shapes in the images and analyzed statistically. The analysis results for these actual 2,836 images prove that our latest LPR system accepts the attitude angles within its product specifications with enough margins.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Statistical, Syntactic and Structural Pattern Recognition
Abstract: Rule induction algorithms such as Ripper, solve a K > 2 class problem by converting it into a sequence of K-1 two-class problems. As a usual heuristic, the classes are fed into the algorithm in the order of increasing prior probabilities. In this paper, we propose two algorithms to improve this heuristic. The first algorithm starts with the ordering the heuristic provides and searches for better orderings by swapping consecutive classes. The second algorithm transforms the ordering search problem into an optimization problem and uses the solution of the optimization problem to extract the optimal ordering. We compared our algorithms with the original Ripper on 8 datasets from UCI repository [2]. Simulation results show that our algorithms produce rulesets that are significantly better than those produced by Ripper proper.

Keywords: Biometrics, Pattern Recognition for Surveillance and Security, Image and Video Processing
Abstract: Recently, a considerable amount of research has been focused on inverse biometrics, that is, regenerat- ing the original biometric sample from its template. In this work, the first reconstruction approach to recover hand geometry samples from their feature vectors is proposed. Experiments are carried out on the publicly available GPDS Hand DB, where the method has shown a remarkable performance, after reconstructing a very high percentage of the hands included in the dataset. Furthermore, the proposed technique is general, being able to successfully reproduce the original hand shape sample regardless of the information and format of the template used.

Keywords: Pattern Recognition for Search, Retrieval and Visualization, 2D/3D Object Detection and Recognition
Abstract: In this paper, we propose a new scoring method for local feature-based image retrieval. The proposed score is based on the ratio of the probability density function of an object model to that of background model, which is efficiently calculated via nearest neighbor density estimation. The proposed method has the following desirable properties: (1) a sound theoretical basis, (2) effectiveness than IDF scoring, (3) applicability not only to quantized descriptors but also to raw descriptors, and (4) ease and efficiency of calculation and updating. We show the effectiveness of the proposed method empirically by applying it to a bag-of-visual words-based framework and a k-NN voting framework.

Keywords: Pattern Recognition for Bioinformatics, Segmentation, Color and Texture
Abstract: A new machine vision method is presented to detect micro tubes with colored solutions or undissolved compounds in solvents on screening plates used in drug discovery. The method presented herein takes an image of a 96 tubes micro tube rack as input. After applying edge detection on the input image, the circles that characterize the borders of the micro tubes are detected via shape matching. A gradient search is used to extract the colors from the micro tubes. Automatic calibration is applied to adapt to changing light conditions. Experiments with more than two thousand micro tube racks (approx. 170,000 compounds) compiled over three years showed that the method is fast and robust.

Keywords: Machine Learning and Data Mining, Feature Reduction and Manifold Learning, Classification and Clustering
Abstract: Semi-supervised learning is important when labeled data are scarce. In this paper, we develop a novel semi-supervised spectral feature selection technique using label regression and by using ell_{1}-norm regularized models for subset selection. Specifically, we propose a new two-step spectral regression technique for semi-supervised feature selection. In the first step, we use label propagation and label regression 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 lower-dimensional 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 standard face data-sets.

Keywords: Statistical, Syntactic and Structural Pattern Recognition, Classification and Clustering, Machine Learning and Data Mining
Abstract: We present a method to perform graph matching in which the human can interact and impose part of the graph labelling. Humans are very good at finding the correspondences between parts of two images but finding these correspondences is one of the most difficult tasks in pattern recognition. Through simple actions such as impose a node labelling or consider a node labelling is not correct; the user helps the automatic graph-matching algorithm to decide a labelling that is closer to the one that the human desires. This interaction is done through the modification of the initial pairwise costs. The method is independent on the graph-matching algorithm. Practical evaluation, in which the Graduated Assignment has been used, shows that with few interactions, the algorithm arrives at the ideal labelling.

Keywords: Pattern Recognition for Bioinformatics, Enhancement, Restoration and Filtering, Classification and Clustering
Abstract: In this paper, we introduce Waveform Length (WL), a new feature for ElectroEncephaloGraphy (EEG) signal classification which measures the signal complexity. We also propose the Waveformlength Optimal Spatial Filter (WOSF), an optimal spatial filter to classify EEG signals based on WL features. Evaluations on 15 subjects suggested that WOSF with WL features provide performances that are competitive with that of Common Spatial Patterns (CSP) with Band Power (BP) features, CSP being the optimal spatial filter for BP features. More interestingly, our results suggested that combining WOSF with CSP features leads to classification performances that are significantly better than that of CSP alone (80% versus 77% average accuracy respectively).

Keywords: Pattern Recognition for Surveillance and Security, Gesture and Behavior Analysis, Classification and Clustering
Abstract: In this paper, we propose a novel unsupervised online learning trajectory analysis method based on weighted directed graph. Each trajectory can be represented as a sequence of key points. In the training stage, unsupervised expectation-maximization algorithm (EM) is applied for training data to cluster key points. Each class is a Gaussian distribution. It is considered as a node of the graph. According to the classification of key points, we can build a weighted directed graph to represent the trajectory network in the scene. Each path is a category of trajectories. In the test stage, we adopt online EM algorithm to classify trajectories and update the graph. In the experiments, we test our approach and obtain a good performance compared with state-of-the-art approaches.

Keywords: Machine Learning and Data Mining, Classification and Clustering
Abstract: Class imbalance limits the performance of most learning algorithms since they cannot cope with large differences between the number of samples in each class, resulting in a low predictive accuracy over the minority ones. Several algorithms achieving more balanced performance in case of binary learning have been proposed, while few researches exists in case of multiclass learning. This paper proposes a new reconstruction rule for the One-per-Class (OpC) decomposition method that, distinguishing between safe and dangerous classifications using sample classification reliability, compensates class imbalance in multiclass recognition problems and reduces effects due to the skewness between classes. The approach has been successfully tested on five datasets using three different classification architectures, and it favourably compares with results provided both by a multiclass classifier and by a popular OpC reconstruction rule.

Keywords: 2D/3D Object Detection and Recognition, Vision for Robotics, Detection, Separation and Segmentation
Abstract: The detection of moving objects for the surveillance and monitoring has been studied in the computer vision community for many years. Traditionally, the studies assume the use of a stationary camera. When using a 3D point cloud, research is restricted to the fixed laser scanner because of the slow data acquisition time. In this paper, we propose a method for detecting moving objects based on a freely moving sensor that provides two-dimensional-three-dimensional(2D-3D) fused data. Our method is a frame-differencing, which compares two consecutive frames combining visual features and a 3D point cloud. The RANSAC and ICP algorithms are applied for more accurate results. The moving objects can be separated in the 3D point cloud by adopting RANSAC outliers.

Keywords: Vision for Graphics
Abstract: In reconstructing 3-D shape from images based on feature points, we usually define a triangular mesh that has those feature points as vertices, and display the object as a polyhedron. If the object itself is a polyhedron, however, some of the displayed edges may be inconsistent with the true shape. For this problem, Nakatsuji et al. proposed a method that automatically eliminates such inconsistencies by flipping triangular edges using a special template. In this paper, we point out a problem of the Nakatsuji's method and propose a new method for overcoming this problem. Using real images, we demonstrate that our method effectively works for complicated mash inconsistencies.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Surveillance and Security, Image and Video Understanding
Abstract: Abstract This paper examines a new problem in large scale stream data: abnormality detection which is localised to a data segmentation process. Unlike traditional abnormality detection methods which typically build one unified model across data stream, we propose that buliding multiple detection models focused on different coherent sections of the video stream would result in better detection performance. One key challenge is to segment the data into coherent sections as the number of segments is not known in advance and can vary greatly across cameras; and a principled way approach is required. To this end, we first employ the recently proposed infinite HMM and collapsed Gibbs inference to automatically infer data segmentation followed by constructing abnormality detection models which are localised to each segmentation. Importantly, the use of this non-parametric method means that the number of models do not need to be specified apriori and can grow with the data. We demonstrate the superior performance of the proposed framework in a real-world surveillance camera data over 14 days.

Keywords: Features and Image Descriptors, Segmentation, Color and Texture, Statistical, Syntactic and Structural Pattern Recognition
Abstract: This paper proposes a new feature extraction technique for indexing and matching historical images. The complexity of these historical images troubles the existing indexing approaches due to their line patterns. Our indexing method relies on the global segmentation integrated with the knowledge of edge density to deal with the line patterns and eliminate over-segmented regions. Then the historical images are represented by a set of ROI descriptors and the retrieval is performed by matching sorted features. Experiments and system evaluations with and without ground truth are presented and discussed.

Keywords: Segmentation, Color and Texture, 2D/3D Object Detection and Recognition, Scene Understanding
Abstract: Gestalt principles have been studied for about a century and were used for various computer vision approaches during the last decades, but became unpopular because the many heuristics employed proved inadequate for many real world scenarios. We show a new methodology to learn relations inferred from Gestalt principles and an application to segment unknown objects, even if objects are stacked or jumbled and tackle also the problem of segmenting partially occluded objects. The relevance of the relations for object segmentation is learned with support vector machines (SVMs) during a training period. We present an evaluation of the relations and show results at the end.

Keywords: Segmentation, Color and Texture, Image and Video Processing, Character and Text Recognition
Abstract: In this paper a new connectivity model is introduced which allows combined clustering and partitioning of structures without distortion, in contrast to mask connectivity. An algorithm to compute morphological attribute filters based on Max-Trees for this new form of connectivity is presented. It is shown that the new form of connectivity is effective in clustering diacritics together with the appropriate letters in historical documents, whilst separating letters from different lines in a single Max-Tree algorithm.

Keywords: Motion, Tracking and Video Analysis
Abstract: We may represent human actions as a bag of spatio-temporal visual words extracted from input video sequences. For human action categorization, labeled LDA (L-LDA) is an extension of latent Dirichlet allocation (LDA) by providing action class labels to each video. To handle parameter uncertainty in L-LDA, this paper further extends L-LDA within the type-2 fuzzy set (T2 FS) framework, referred to as T2 L-LDA. Because the membership function (MF) of T2 FS is three-dimensional, we can use the primary MF to measures the degree of uncertainty that a visual word belongs to a specified human action category, and use the secondary MF to evaluate the fuzziness of the primary MF itself. We also develop the T2 fuzzy belief propagation (T2 BP) algorithm for approximate inference and parameter estimation based on T2 FS operations. On the KTH human motion data set, our results show that T2 L-LDA is able to enhance the overall performance in human action categorization.

Keywords: 2D/3D Object Detection and Recognition, Classification and Clustering, Mixed and Augmented Reality
Abstract: In many applications which require real-time keypoint recognition such as Augmented Reality, Random Ferns (RF) is widely used due to its runtime performance. It relies on an offline training phase during which runtime computational burdens are delegated. This leads to robust, accurate, and frame-rate performance. However, it requires significant amounts of memory, and this has been an obstacle to its use in industry, especially in mobile environments. In this paper, we propose Lightweight Random Ferns to reduce the memory requirements of RF by modifying the representation of probabilities used in ferns to a single bit from floating point. As a result, the total memory requirements of RF are significantly reduced.

Keywords: Scene Understanding, Machine Learning and Data Mining
Abstract: In this paper, an automatic image annotation (AIA) method using Gaussian mixture model (GMM) is discussed. Supervised multiclass labeling (SML), which is a notable AIA method using GMM, has a problem of low annotation performances of labels that have a few training samples because of over fitting. In the present study, we propose to introduce a cross entropy based constraint into SML. According to the proposed method, while probabilistic models of labels are trained independently as is the case with SML, the optimization of whole probabilistic models is achieved, and therefore over fitting is suppressed. As the result of extensive evaluation tests, the proposed method obtained the best annotation performance in existing parametric methods of AIA.

Keywords: Motion, Tracking and Video Analysis
Abstract: In this paper we present an effective and fast track- ing algorithm, in which object tracking is achieved by solving L2-regularized least square (L2-RLS) problem- s within a Bayesian inference framework. Firstly, we model the appearance of the tracked target with P- CA basis vectors and square templates which make the tracker not only exploit the strength of subspace repre- senation but also explicitly take partial occlusion into consideration. Secondly, we adopt the L2-regularized least square method to solve the proposed representa- tion model. Compared with the complicated L1-based algorithm, it provides a very fast performance without loss of accuracy in handling the tracking problem. In addition, a novel likelihood function and a refined up- date scheme are designed to further improve the ro- bustness of our tracker. Both qualitative and quantita- tive evaluations on challenging videos demonstrate that the proposed method performs favorably against sever- al state-of-the-art tracking algorithms.

Keywords: Computational Photography, Stereo and Image-Based Modeling, Image and Video Processing
Abstract: We propose a new direct microlens array imaging system without additional refractive optics. A diffuse led light source and a hot embossed micro lens array enable a compact and very low cost structure for portable devices. A light source aperture acts as a field stop for microlenses, and separate micro apertures are not required by individual lenses. The area scalability achieved with polymeric lenslet arrays enables fast screening of large sample areas by removing the need for spatial scanning. A light field rendering method based on mosaicing was developed, and the performance of polymer microlenses is compared to glass microlenses with the suggested system. The utilized polycarbonate lenses enabled 8.8 m resolution.

Keywords: 2D/3D Object Detection and Recognition, Statistical, Syntactic and Structural Pattern Recognition, Pattern Recognition for Bioinformatics
Abstract: In this paper a novel 2D shape recognition approach is proposed. The main idea is to exploit in this context the huge amount of work carried out by bioinformaticians in the biological sequence analysis research field. In the proposed approach, we encode shapes as biological sequences, employing standard and well established sequence alignment tools to devise a similarity score, finally used in a nearest neighbour scenario. Despite its simplicity, obtained results on standard datasets are really encouraging.

Keywords: Features and Image Descriptors, Image and Video Processing, Low-Level Vision
Abstract: This paper proposes a method for constructing local image descriptors which efficiently encode texture information and are suitable for histogram based representation of image regions. The method computes a binary code for each pixel by linearly projecting local image patches onto a subspace, whose basis vectors are learnt from natural images via independent component analysis, and by binarizing the coordinates in this basis via thresholding. The length of the binary code string is determined by the number of basis vectors. Image regions can be conveniently represented by histograms of pixels' binary codes. Our method is inspired by other descriptors which produce binary codes, such as local binary pattern and local phase quantization. However, instead of heuristic code constructions, the proposed approach is based on statistics of natural images and this improves its modeling capacity. The experimental results show that our method improves accuracy in texture recognition tasks compared to the state-of-the-art.

Keywords: Features and Image Descriptors, Segmentation, Color and Texture, 2D/3D Object Detection and Recognition
Abstract: This paper investigates the segmentation of multiple planar surfaces from 3D point clouds. A Principle Component Analysis (PCA) based covariance technique is used for segmentation which is one of the most popular approaches in point cloud processing. It is well known that PCA is very sensitive to outliers and does not give reliable estimates for segmentation. We propose a statistically robust segmentation algorithm using a fast-minimum covariance determinant based robust PCA approach to get the local covariance statistics. This results in more reliable, robust and accurate segmentation. The application of the proposed method to simulated and terrestrial laser scanning point cloud datasets gives good results for multiple planar surface extraction and shows significantly better performance than PCA based methods. The algorithm has the potential for non-planar complex surface reconstruction.

Keywords: 2D/3D Object Detection and Recognition, Pattern Recognition for Surveillance and Security, Pattern Recognition for Search, Retrieval and Visualization
Abstract: Person re-identification is an important problem in visual surveillance where appearance plays a key role. Color is one of the widely used appearance features and utilizing more color spaces doesn�t imply benefit of performance enhancement. That�s because the poor performance color spaces influence on the high ones. So it is significant to evaluate the performance of different color spaces for person re-identification. In this paper, we propose a novel approach, called as random ensemble of color features (RECF), where we build a random forest to learn the similarity function of pairs of person images using color features from 6 kinds of popular color spaces (RGB, normalized RGB (NRGB), HSV, YCbCr, CIE XYZ and CIE Lab). We carry out experiments on the challenging dataset VIPeR to show the performances of different color spaces and their combination. We find out that the combination of NRGB, HSV, YCbCr and CIE Lab color spaces achieves the best performance, and our approach alleviates the over-fitting problem when there are limited training data.

Keywords: 2D/3D Object Detection and Recognition, Features and Image Descriptors, Vision for Robotics
Abstract: We propose a new framework for recognizing three-dimensional (3D) objects that needs only a few reference images. We tackle one issue with 3D object recognition; the number of reference images must be traded-off against recognition accuracy. We assume that a reference image is a photo of the object from an arbitrary position, typical of those used to exhibit goods on the web. The framework, first, estimates a 3D model of the object from a video sequence (the input), and generates a single target image by spherical projection of the 3D model and texture. The framework recognizes the object by matching the target image against the reference images stored in the database. The descriptor matching approach is used, with the image getting the most votes becoming the recognition result. The vote score is computed from the vector distance between key-points, not only for a given key-point but also for key-points in its vicinity.

Keywords: 2D/3D Object Detection and Recognition, Remote Sensing, Scene Understanding
Abstract: In this paper we present a new model for joint extraction of vehicles and coherent vehicle groups in airborne LIDAR point clouds collected from crowded urban areas. Firstly, the 3D point set is segmented into terrain, vehicle, roof, vegetation and clutter classes. Then the points with the corresponding class labels and intensity values are projected to the ground plane, where the optimal vehicle and traffic segment configuration is described by a Two-Level Marked Point Process (L2MPP) model of 2D rectangles. Finally, a stochastic algorithm is utilized to find the optimal configuration.

Keywords: Machine Learning and Data Mining, Classification and Clustering, Vision for Robotics
Abstract: During training and classification, instances are drawn from the instance space and mapped to the feature space. We focus on the problem of detecting hidden changes in the functions that map instances to feature vectors during classification. We call such changes feature shift and introduce an on-line method for detecting it. Our method is based on a robust similarity measure that uses one-class SVM to monitor distributional changes in the feature space. Unlike previous methods, ours can distinguish between changes in priors and feature shift. The method is empirically evaluated on visual categorization tasks and its advantage verified.

Keywords: Motion, Tracking and Video Analysis, Pattern Recognition for Surveillance and Security, Classification and Clustering
Abstract: This paper introduces CMV100, a new research dataset for people tracking and re-identification in sparse camera networks. Baseline methods for re-identification performance analysis are also proposed. The dataset consist of over 400 indoor video sequences in total. The number of visually distinctive human objects is 100, and each person appears in three different views on average and in five at maximum. The dataset evaluation is performed using sequence matching methods based on adaptive boosting and CART decision trees. The preliminary experiments show the challenging characteristics of the new dataset and serve as a practical starting point for future improvements.

Keywords: Vision for Graphics, Mixed and Augmented Reality, Segmentation, Color and Texture
Abstract: Capturing appearance of material with respect to illumination and viewing directions is crucial to achieve realistic visual experience in virtual environments. The capturing process is time demanding or requires a specific shape of the captured material. Therefore, we propose a method of such a data reconstruction from very sparse measurements, whose placement allows for continuous and fast acquisition, from which can benefit future acquisition setups. The proposed approach was tested on a number of view- and illumination- dependent samples and showed a promising performance in terms of whole data-space reconstruction speed and visual quality.

Keywords: Features and Image Descriptors, 2D/3D Object Detection and Recognition, Motion, Tracking and Video Analysis
Abstract: In this paper we propose a new 3D descriptor for people re-identification using depth cameras. The descriptor is based on a 3D cylindrical grid that stores color variations with angle and height. A robust comparison method is also presented. This method ignores variations in brightness and contrast of descriptors obtained at possibly different locations. However it retains the ability to distinguish between different colors (e.g red vs. blue) or luminance patterns. Extensive results on three different scenarios are presented and prove the validity of the approach. The data sets used have been publicly released.

Keywords: Low-Level Vision, Statistical, Syntactic and Structural Pattern Recognition, Segmentation, Color and Texture
Abstract: Scale invariance is a desirable property for many vision tasks such as image segmentation and classification. One way to achieve such invariance is to collect images containing objects of all scales and then train a classifier. In practice, however, only a finite number of images at a finite number of scales can be collected, and this poses the problem of scale sampling. In this paper, we focus on how to properly sample over scales in order to solve scale-invariant image segmentation. The ideal distributions of images and features in a scale-invariant setting are derived, and their implications for scale sampling and feature extraction are studied. Some basic image segmentation experiments are conducted to examine the sampling rules proposed, which show that it is possible to train a scale-invariant classifier from a single image.

Keywords: 2D/3D Object Detection and Recognition, Image and Video Processing
Abstract: This paper proposes a new edge descriptor named centripetal-SIFT edge descriptor. A method for symmetric object detection is presented based on symmetry and centripetal-SIFT descriptor in real unsegmented images. Our method includes three main steps: 1) The dominant symmetry axis is located based on SIFT feature point; 2) Image edge points are extracted in scale space and described by the proposed centripetal-SIFT descriptor; 3) Symmetric objects are located as the result of computing symmetric edge points. The superior performance of our method is demonstrated by experimental results under complex backgrounds.

Keywords: Cognitive and Embodied Vision, Low-Level Vision, Image and Video Processing
Abstract: In this paper, we propose a new region-based saliency model to simulate the human visual attention. First, we construct a pixel-level fully-connected graph representation for an image, and perform normalized cut to segment the image based on the proximity and similarity principles. After obtaining image regions, we reconstruct a region-based fully-connected graph. Based on the saliency principle ``center-surround contrast'', we define new dissimilarity functions in terms of several visual features. Finally we run a random walk on the region graph and apply site entropy rate to measure the region saliency. We evaluate the proposed model on a public dataset consisting of 120 images. Experimental results demonstrate that our model predicts eye fixations more accurately than the other four state-of-the-art methods. We also apply our saliency model to improve the performance of image retargeting.

Keywords: Motion, Tracking and Video Analysis, Human Computer Interaction, Gesture and Behavior Analysis
Abstract: Human vision can perceive body movements and actions effortlessly. In contrast, it is still a very challenging task for machines to have comparable performance. Many research results have shown that both visual attention and perceptual organization are crucial for visual perception tasks. In recent years, gesture recognition for HCI has drawn more attention because of high application demands. Based on visual perceptual theories and hypotheses, we propose a 3D gesture recognition framework in a coherent and biologically plausible manner. It mainly includes perceptual gesture feature extraction, hierarchical salience map construction and qualitative reasoning for gesture recognition.

Keywords: 2D/3D Object Detection and Recognition
Abstract: Almost all of the existing research on 3D face recognition is based on static 3D images. 3D videos are believed to provide more information in terms of both the shape as well as the dynamics of an individual's face. This paper presents a system which exploits the spatiotemporal information in 3D videos for the task of face recognition. An algorithm for automatic normalization of raw 3D videos is also given. After the detection of the nose tip, all meshes of the 3D video are cropped and uniformly sampled to form range videos. Spatiotemporal Local Binary Pattern (LBP) descriptors are used for feature extraction from the range videos. For classification, a linear multiclass Support Vector Machine (SVM) is used. The system is trained on videos of a person with different facial expressions and tested on a video with new facial expression. Experimental results on the largest currently available 3D video database, BU_4DFE, show a high recognition rate of 92.68%.

Keywords: Features and Image Descriptors, Classification and Clustering, Image and Video Understanding
Abstract: In this paper we present a new method for object categorization. Firstly an image representation is obtained by the proposed hierarchical learning method consisting of alternating between local coding and maximum pooling operations, where the local coding operation induces discrimination while the image descriptor and maximum pooling operation induces invariance in hierarchical architecture. Then the obtained effective image representation is passed to a linear classifier which is suitable for large databases for object categorization. We have demonstrated that the proposed method is robust to image variations and has low sample complexity.

Keywords: Segmentation, Color and Texture, Features and Image Descriptors, Low-Level Vision
Abstract: Center-surround measurements are widely used for saliency detection but with some disadvantages: 1) Center-surround operation may cause inaccurate segmentation and even involve incorrect detection results; 2) In most situations, only using center-surround feature is not efficient to encode object saliency. To overcome these disadvantages, we describe a novel measurement, namely Corner-Surround Contrast (CSC), to segment salient regions from backgrounds. To explore the effects of CSC feature, a kernel-based fusing framework is designed to produce the saliency map automatically and infer the binary segmentation using graph cut algorithm. The experiments demonstrate the promising performance of our method in terms of segmentation accuracy and saliency localization.

Keywords: Segmentation, Color and Texture, Detection, Separation and Segmentation
Abstract: This paper proposes a functional that assigns low `energy' to sets of subsets of the image domain consisting of a number of possibly overlapping near-circular regions of approximately a given radius: a `gas of circles'. The model can be used as a prior for object extraction whenever the objects conform to the `gas of circles' geometry, eg cells in biological images. Configurations are represented by a multi-layer phase field. Each layer has an associated function, regions being defined by thresholding. Intra-layer interactions assign low energy to configurations consisting of non-overlapping near-circular regions, while overlapping regions are represented in separate layers. Inter-layer interactions penalize overlaps. Here we present a theoretical and experimental analysis of the model.

Keywords: Features and Image Descriptors, Pattern Recognition for Search, Retrieval and Visualization, Classification and Clustering
Abstract: Recent years have seen an increasing interest in codebook-based model(bag-of-words--BoW) for image representation, which includes the basic bag-of-words model and its improved version for local descriptor reconstruction with sparse coding (SC) and locality-constrained linear coding (LLC) etc.. Although the recent coding strategies in the BOW model can lead to prospect performance using large amounts of codes (codebooks) for image classification, it is usually computational expensive for obtaining the global image representation through calculating the similarities between each local descriptor and all codes. Therefore, this study proposes to represent a local descriptor with an adaptive code or its variation modes (adaptive sub-domain) in a small set of codebooks. The proposed strategy can adaptively select one code to saliently representation, or adaptively select one sub-domain of a code for group sparse reconstruction of a local descriptor. Due to computational cost mainly on the similarity calculation between local descriptors and the predefined codebooks, our proposed strategy using small set of codebook can greatly reduce computational time, and in addition, shows prospect performances for image classification on an scene database, called OMRON scene dataset, and the benchmark data: 15 natural scene dataset.

Keywords: Physics-Based Vision, Low-Level Vision, Image and Video Processing
Abstract: This paper proposes a daylight spectrum model under weather conditions from clear sky to cloudy. Various color constancy theories have been proposed to estimate the spectral properties of the illumination and objects in a scene from color information obtained by a camera. A finite dimensional linear model is used to model the spectral properties. When the range of wavelength obtained by a hyperspectral camera is wide, e.g. 300nm-1000nm, the wavelength range of the conventional basis vectors of daylight (300nm-830nm) is insufficient. Bird Model is a daylight spectrum model with a wide range of wavelength from 300nm to 4000nm. However, it is not practical since it can be applied only under clear sky. In this study, this restriction in weather is solved by taking the behavior of daylight passing through clouds into consideration. When daylight passes through clouds, spectral property of the daylight does not change and only intensity of the daylight changes. Based on the characteristic, a daylight spectrum is simply represented as a linear combination of daylight components under clear sky with one parameter describing the degree of cloudiness in the proposed model. Experimental results show that the proposed model is promising.

Keywords: Vision for Robotics, Geometric and Photometric Registration, Low-Level Vision
Abstract: In this paper we propose an extrinsic calibration method of regular camera and single line scanning lidar which are widely utilized together. Based on a long ignored aspect that infra-red (IR) source for commonly used lidars lies in the response range of regular cameras, we employ an auxiliary IR filter (blocking natural light and letting IR pass) in order for camera to image the scan traces of lidar by prolonging exposure time. Then by scanning V-shaped target such as the intersection of two smooth walls, corresponding lines (or points) on scan plane and image are found through line fitting. With these high confident correspondences, extrinsic parameters (known camera) or a planar homography (unknown camera) can easily be calculated. Furthermore, two evaluation methods, namely line alignment error and two-view point alignment error, are developed. Experiments show that our method greatly simplifies the calibration procedure and outperforms the state-of-the-art in accuracy by using only one-tenth of the calibration data.

Keywords: 2D/3D Object Detection and Recognition, Motion, Tracking and Video Analysis, Vision for Graphics
Abstract: This paper proposes a novel technique for 3-D recovery of a non-rigid object, such as a human in motion, from a single camera view. To achieve the 3-D recovery, the proposed technique performs segmentation of an object under deformation into respective parts which are regarded as rigid. For high accuracy segmentation, multi-stage learning and local subspace affinity are employed in this stage. Each part recovers its 3-D shape by applying the factorization method to it. This is the initial solution. The shape of each part is then refined by applying a quadratic model to the initial solution. The entire 3-D recovery of the object is finally performed using the common points among the segmented parts. The experiments employing a synthetic non-rigid object and real human motion data show effectiveness of the proposed technique.

Keywords: Motion, Tracking and Video Analysis, Machine Learning and Data Mining, Pattern Recognition for Surveillance and Security
Abstract: Recently many appearance based visual tracking algorithms have been investigated, aimed at building robust appearance models against challenges brought by the varying appearance of the target as well as the unconstrained environment. More often adaptive appearance models were used to capture these variances over time, but this may sometimes result in losing the target (drifting) due to inappropriate update of the model. In this paper an online form of Random Ferns classifier is proposed to accomplish the task of robust appearance modeling with a constrained updating strategy against the potential incorrect update induced by runtime noise. Experiments on challenging benchmark video sequences have been conducted and improvement is observed when compared with recent state-of-the-art algorithms.

Keywords: Motion, Tracking and Video Analysis, Vision for Robotics, 2D/3D Object Detection and Recognition
Abstract: In order to track and estimate the pose of known rigid objects with high accuracy in unconstrained environment with light disturbance, scale changes and occlusion, we propose to combine 3D particle filter (PF) framework with algebraic pose optimization in a closed loop. A new PF observation model based on line similarity in 3D space is devised and the output of 3D PF tracking, namely line correspondences (model edges and image line segments), are provided for algebraic line-based pose optimization. As a feedback, the optimized pose serves as a particle with high weight during re-sampling. To speed up the algorithm, a dynamic ROI is used for reducing the line detection and search space. Experiments show our proposed algorithm can effectively track and accurately estimate the pose of freely moving 3D rigid objects in complex environment.

Keywords: Motion, Tracking and Video Analysis, Classification and Clustering, Image and Video Understanding
Abstract: This paper introduces a novel video presentation term spatial-temporal pyramid sparse coding (STPSC) which characterizes both the spatial and temporal aspects of the video. Specifically, the co-occurrences of visual words are computed with respect to the spatial layout and the sequencing of the features in the video. The representation captures both the spatial arrangement and the temporal relationship of the words. Our representation is motivated by the technology spatial pyramid matching (SPM) which is used to recognize scenes in the image. We extend SPM to video analysis combining with sparse coding. Firstly, dense feature points are extracted and represented by displacement information from a dense optical flow field. Then sparse coding is used to quantize the feature descriptors, and the spatial-temporal pyramid is introduced to represent an action. Finally, we use SVM to classify the videos. Experimental results showed improvements over the state-of-the-art techniques on the public action dataset.

Keywords: Scene Understanding, 2D/3D Object Detection and Recognition, Features and Image Descriptors
Abstract: The locality and sparsity constrained encoding methods have shown the good image classification performance in recent papers. Among these methods, the common strategy is encoding one descriptor into one code by a learned codebook and then applying SPM and Pooling strategy to get the final image representation. However, the ignorance of local spatial context has been a barrier to improve their discriminative power. To address this problem, we propose the so called Adjacent Coding (AC), which employs the adjacency of one descriptor to express the local spatial context. Different from traditional coding methods, Adjacent Coding encodes one descriptor and its adjacent neighbors together. In this paper, we further show that AC also keeps the properties of locality and sparsity. Finally, our experiments on the standard benchmarks (Scene 15 and Caltech 101) show our method can outperform the state-of-the-art feature coding methods.

Keywords: Motion, Tracking and Video Analysis, Feature Reduction and Manifold Learning, Pattern Recognition for Surveillance and Security
Abstract: This paper addresses issues of online learning and occlusion handling in video object tracking. Although manifold tracking is promising, large pose changes and long-term partial occlusions of video objects remain challenging. We propose a novel manifold tracking scheme that tackles such problems, with the following main novelties: (a) Online estimation of object appearances on Grassmann manifolds; (b) Optimal criterion-based occlusion handling during online learning; (c) Nonlinear dynamic model for appearance basis matrix and its velocity; (b) Bayesian formulations separately for the tracking and the online learning process. Two particle filters are employed: one is on the manifold for generating appearance particles and another on the linear space for generating affine box particles. Tracking and online updating are performed in alternative fashion to mitigate the tracking drift. Experiments on videos have shown robust tracking performance especially when objects contain significant pose changes accompanied with long-term partial occlusions. Evaluations and comparisons with two existing methods provide further support to the proposed method.