MaxHash for Fast Face Recognition and Retrieval

This paper presents a fast method for recognition and retrieval of face images in large face datasets. In this approach, we generate binary hash codes for deep face features extracted using FaceNet model. Unlike the distance of real valued feature vectors, the distance in the Hamming space is computed with a simple XOR operation. We generate a candidate set using the Hamming distance metric, then a fine grained search is used to find the corresponding face picture in this small set. Experiments on LFW dataset show that for 64 bit hash codes an image of the corresponding person is always in the candidates set of only 48 items.