ECCV 2012 - LNCS 7572-7578 and 7583-7585

Adaptive Registration for Occlusion Robust 3D Face Recognition

Nese Alyuz¹, Berk Gokberk², and Lale Akarun¹

¹Department of Computer Engineering, Boaziçi University, Istanbul, Turkey
nese.alyuz@boun.edu.tr
akarun@boun.edu.tr

²Department of Electrical Engineering Mathematics and Computer Science, University of Twente, The Netherlands
b.gokberk@utwente.nl

Abstract. Occlusions over facial surfaces cause performance degradation for face registration and recognition systems. In this work, we propose an occlusion-resistant three-dimensional face registration method. First, the nose area is detected on a probe face using curvedness-weighted convex shape index map. Then, probable eye and mouth patches are detected and checked for validity. An adaptive model is constructed by selecting valid patches of the average face model. Finally, registration is handled with the Iterative Closest Point algorithm, where the adaptive model is used as the reference. The UMB-DB face database is used to evaluate the registration system: The nose detector has 100% and 93.90% accuracy, for the non-occluded and occluded images, respectively. A simple global depth-based recognition experiment is done to evaluate the registration performance: Our adaptive model-based registration scheme improves rank-1 recognition rate by 16%, when compared with the nose-based alignment approach.

Keywords: 3D face registration, regional face registration, face registration under occlusion

LNCS 7585, p. 557 ff.

Full article in PDF | BibTeX