Deep Supervised Hashing with Pairwise Bit Loss
Jiabao Wang, Yang Li, Xiancai Zhang, Zhuang Miao, Gang Tao · 2017
Low dimensional binary hashing is the key point in large-scale image retrieval and person re-identification (re-ID). To promote the performance, we explore the possibility of deep supervised hashing using the label information. A pairwise bit loss is proposed to measure the difference between two features extracted from two intra-class images by CNNs. Furthermore, a Siamese network architecture is proposed to conduct deep feature hashing learning by combining the classification loss and the proposed pairwise bit loss. The input of our proposed network has only positive pair, without the negative pair. Experiments show that our method can outperform other methods to achieve the state-of-the-art performance in image retrieval and person re-ID benchmarks.