Patching Your Clothes: Semantic-Aware Learning for Cloth-Changed Person Re-Identification

Great advances have been observed in conventional person re-identification (Re-ID), which heavily relies on the assumption that the cloth remains unchanged. However, this dramatically limits their applicability in practical cloth-changed scenarios, leading to dramatic performance drop. Existing cloth-changed methods mainly exploit the body shape information, ignoring the relation between different clothes of the same identity. In this paper, we present a powerful semantic-aware patching strategy for clothes augmentation. It greatly enriches the cloth styles by randomly assembling the semantic cloth patches, simulating the appearances of the same person under different clothes. This augmentation strategy has two primary advantages: 1) It significantly reinforces the robustness against clothes variations without additional cloth collection. 2) It does not damage semantic structure, fitting well with cloth-unchanged scenarios. To further address the uncertainty in cloth changed, a Semantic Part-aware Feature Learning scheme is incorporated to mine fine-grained granularities, addressing the misalignment issue under changed clothes. Extensive experiments conducted on both clothing-changed and cloth-unchanged tasks demonstrate our proposed method's superiority, consistently improving the performance over various baselines.