Style Normalization and Restitution for Generalizable Person Re-identification

Existing fully-supervised person re-identification (ReID) methods usuallysuffer from poor generalization capability caused by domain gaps. The key tosolving this problem lies in filtering out identity-irrelevant interference andlearning domain-invariant person representations. In this paper, we aim todesign a generalizable person ReID framework which trains a model on sourcedomains yet is able to generalize/perform well on target domains. To achievethis goal, we propose a simple yet effective Style Normalization andRestitution (SNR) module. Specifically, we filter out style variations (e.g.,illumination, color contrast) by Instance Normalization (IN). However, such aprocess inevitably removes discriminative information. We propose to distillidentity-relevant feature from the removed information and restitute it to thenetwork to ensure high discrimination. For better disentanglement, we enforce adual causal loss constraint in SNR to encourage the separation ofidentity-relevant features and identity-irrelevant features. Extensiveexperiments demonstrate the strong generalization capability of our framework.Our models empowered by the SNR modules significantly outperform thestate-of-the-art domain generalization approaches on multiple widely-usedperson ReID benchmarks, and also show superiority on unsupervised domainadaptation.