Semantic-guided hashing learning for domain adaptive retrieval

Recently, domain adaptive retrieval has aroused much attention. However, most existing methods are proposed under the single-domain assumption. They neglect two issues: a) the data distribution discrepancy between the retrieved set (source domain) and query set (target domain); and b) the semantic discrepancy between the features and labels. In this work, we propose a novel transferable hashing method to address these two issues, termed Semantic-Guided Hashing Learning (SGHL). First, the marginal and conditional distributions between the source and target domains are aligned to reduce the distribution discrepancy between the two domains. Then, we embed the semantic information of the source domain into a latent semantic space to alleviate the semantic discrepancy between the features and labels. Moreover, linear embedding is explored with orthogonal transformation to minimize the quantization loss between the latent semantic space and Hamming space. At last, an iterative algorithm is designed to generate hash codes directly. Extensive experiments on four widely-used cross-domain retrieval datasets demonstrate that SGHL outperforms the state-of-art hashing methods.