Deep momentum uncertainty hashing

Combinatorial optimization (CO) has been a hot research topic because of its theoretic and practical importance. As classic CO problem, deep hashing aims to find an optimal code for each data from finite discrete possibilities, while the nature brings big challenge process. Previous methods usually mitigate this by binary approximation, substituting codes real-values via activation functions or regularizations. However, such approximation leads uncertainty between ones, degrading retrieval performance. In paper, we propose novel Deep Momentum Uncertainty Hashing (DMUH). It explicitly estimates during training leverages information guide Specifically, model bit-level measuring discrepancy output network that momentum-updated network. The bit indicates approximate bit. Meanwhile, mean all bits in can be regarded as image-level uncertainty. embodies corresponding input image. image with higher are paid more attention optimization. To best our knowledge, is first work study bits. Extensive experiments conducted on four datasets verify superiority method, including CIFAR-10, NUS-WIDE, MS-COCO, million-scale dataset Clothing1M. Our method achieves performance surpasses existing state-of-the-art large margin.