Deep Features for Recognizing Disguised Faces in the Wild

Unconstrained face verification is a challenging problem owing to variations in pose, illumination, resolution of image, age, etc. This problem becomes even more complex when the subjects are actively trying to deceive face verification systems by wearing a disguise. The problem under consideration here is to identify a subject under disguises and reject impostors trying to look like the subject of interest. In this paper we present a DCNN-based approach for recognizing people under disguises and picking out impostors. We train two different networks on a large dataset comprising of still images and video frames with L2-softmax loss. We fuse features obtained from the two networks and show that the resulting features are effective for discriminating between disguised faces and impostors in the wild. We present results on the recently introduced Disguised Faces in the Wild challenge dataset.