LTReID: Factorizable Feature Generation with Independent Components for Long-Tailed Person Re-Identification

With the rapid increase of large-scale and real-world person datasets, it is crucial to address problem long-tailed data distributions, i.e., head classes have large number images while tail occupy extremely few samples. We observe that imbalanced distribution likely distort overall feature space impair generalization capability trained models. Nevertheless, this has been rarely investigated in previous Re-Identification (ReID) works. In paper, we propose a novel Long-Tailed (LTReID) framework simultaneously alleviate class-imbalance hard-imbalance problems. Specifically, each real decomposed into multiple independent components with two decorrelation losses. Then these are randomly aggre- gated generate more fake features for than ones, resulting class-balance between classes. For hard-balance easy hard samples, utilize adversarial learning ones. The proposed can be an end-to-end manner avoids increasing time complexity inference Moreover, comprehensive experiments conducted on four ReID datasets so as validate effectiveness advantage module. Our results show when either balanced or LTReID achieves superior performance over state-of-the-art methods.