Deep Triplet Supervised Hashing

Hashing is one of the most popular and powerful approximate nearest neighbor search techniques for large-scale image retrieval. Most traditional hashing methods first represent images as off-the-shelf visual features and then produce hash codes in a separate stage. However, off-the-shelf visual features may not be optimally compatible with the hash code learning procedure, which may result in sub-optimal hash codes. Recently, deep hashing methods have been proposed to simultaneously learn image features and hash codes using deep neural networks and have shown superior performance over traditional hashing methods. The current state-of-the-art deep hashing method DPSH [14] has demonstrated great performance but it suffers the problem of unbalanced supervision signals. To address this issue, we propose two novel deep hashing methods DPSH-Weighted and DTSH. At the heart of our methods DPSH-Weighted and DTSH is the novel formulation of weighted pairwise label likelihood and triplet label likelihood. Our methods learn images features and hash codes by maximizing the label likelihood. One common problem that most deep hashing methods face is how to train a network to output binary codes. We provide in-depth analysis of two typical methods to train such a network. Extensive experiments on four public benchmark datasets, including CIFAR-10 [10], NUSWIDE [3], Standford Cars [9] and UKBench [23], show that both DPSH-Weighted and DTSH outperform the state-of-the-art method DPSH [14], and DTSH can obtain the highest performance among all the current deep hashing methods.