Occlusions in Face Recognition: a 3D Approach
Alessandro Colombo, Claudio Cusano, Raimondo Schettini · 2009
The presence of occluding objects in face recognition and more generally in object recognition tasks, is a "far from solved" problem. Here a solution has been presented which is composed of three core modules (detection, normalization and occlusion detection/face restoration) that could be employed in any 3D recognition system in order to improve its robustness. The results are quite promising and indicate that the use of 3D data may simplify the problem. However, there are still a lot of open issues. First of all, is it possible to reduce the errors introduced by the automatic normalization of faces? Secondly, how does the proposed algorithm perform in the presence of emphasized facial expressions? How could the proposed algorithm be integrated with a facial expression tolerant solution? Another important issue regards the comparison between face restoration approaches versus partial matching approaches. It is not clear at present if the missing or covered parts of the face should be reconstructed or simply detected and ignored, letting a partial matching strategy perform recognition. Future studies must take into account all these aspects. A large database containing real occlusions and various types and degrees of facial expressions should be adopted in order to allow an in-depth study of the problem and the proposed solutions.