A Siamese Network for Face Verification with Depth Images

Abstract

Face verification is one of the most researched and demanding tasks in the field of computer vision. The task of Face verification is to check whether two input faces are the same object. With the development of depth cameras and their reliability against light changes, face verification has become more widely used in many scenarios, for example night driving. This paper constructs a deep Siamese architecture for face verification based on two same fully convolutional networks, relying on depth images. Despite the lack of deep-oriented depth-based datasets, the network only relies on the small amount of depth data available on Pandora dataset for training and still achieves state-of-art results and real time performance, and the network also get excellent results during the variation of head pose.