Evaluation of discrimination power of facial parts from 3D point cloud data

Feature selection from facial regions is a well-known approach to increase the performance of 2D image-based face recognition systems. In case of 3D modality, the approach of region-based feature selection for face recognition is relatively new. In this context, this paper presents an approach to evaluate the discrimination power of different regions of a 3D facial surface for its potential use in face recognition systems. We propose the use of weighted average of unit normal vector on the facial surface as the feature for region-based face recognition from 3D point cloud data (PCD). The iterative closest point algorithm is employed for the registration of segmented regions of facial point clouds. A metric based on angular distance between normals is introduced to indicate the similarity between two surfaces of same facial region. Finally, the intra class correlation based discrimination score is formulated to find out the key facial regions such as the eyes, nose, and mouth that are significant while recognizing a person with facial surface PCD.