Face Recognition Using DELF Feature Descriptors on RGB-D Data

Andrey Kuznetsov

Abstract: Face recognition is a very challenging and important task in many areas. This sort of algorithms is used in security systems, for person authorization tasks, person reidentification, etc. In face recognition speed and quality are very important, but in this research, we concentrate on quality improvement. In this paper we propose a new solution for face recognition using RGB-D data obtained with Kinect sensor. The proposed solution is based on a new type of feature descriptors – Deep Learning Features (DELF), which showed high classification results on Google landmarks dataset. We compared DELF descriptors with HOG, MSER and SURF descriptors using two classification schemes: error correcting output codes (ECOC) and decision trees. Conducted experiments showed an improvement in classification quality using F1 score metric when face recognition is based on DELF descriptors and ECOC classifier.