Search

Deep metric learning aims to learn a non-linear function that maps raw-data to a discriminative lower-dimensional embedding space, where semantically similar samples have larger similarity than dissimilar ones. Most existing approaches process each raw-data in two steps, by mapping the raw-data to a higher-dimensional feature space via a fixed backbone, followed by mapping the higher-dimensional feature space to a lower-dimensional embedding space via a linear layer. This paradigm, however, inevitably leads to a Generalization Bottleneck (GB) problem. Specifically, GB refers to a limitation that the generalization capacity of lower-dimensional embedding space is inferior to the higher-dimensional feature space in the test stage. To mitigate the capacity gap between feature space and embedding space, we propose to introduce a fully-learnable module, dubbed Relational Knowledge Preserving (RKP), that improves the generalization capacity of lower-dimensional embedding space by transferring the mutual similarity of instances. Our proposed RKP module can be integrated into a general deep metric learning approach. And, experiments conducted on different benchmarks show that it can significantly improve the performance of original model. [ABSTRACT FROM AUTHOR]

Hashing, a widely-studied tool to the approximate nearest neighbor search, aims to embed samples as compact binary representations. Current approaches to this issue generally seek a low-dimensional Hamming Space where representations are discrete and have smaller intra-class distance and larger inter-class distance. As a result, the performance is often limited by the discrete constraint. In this work, we propose to seek an angularly discriminative Embedding Space where representations are continuous and have smaller intra-class angular margin and larger inter-class angular margin. For our goal is to learn continuous representations rather than discrete hash codes, the problems caused by discrete constraint can be avoided. Besides, in order to further reduce the gap between Embedding Space and Hamming Space, we introduce an additional coordinate-constraint for representations. Our method is simple yet effective. Extensive experiments on the image retrieval task show that it achieves encouraging results on four benchmark datasets. Furthermore, the success of our proposed method demonstrates that leveraging the progress made in representation learning to improve hashing is promising in future. [ABSTRACT FROM AUTHOR]