Clothing Sampling Based on Active Learning For Cloth-Changing Person Re-identification

Yujie Liu, Mingyue Li, Jiansen Jing, Yante Li, Guoying Zhao

Abstract: Cloth-Changing Person Re-Identification (CC-ReID) aims to match the same person with clothing changes. The challenges mainly include two types: same person wearing different clothing and different person wearing similar clothing. The current methods are usually limited by the number and variation of clothing in training data, making it difficult to cope with the latter. To address this issue, this article proposes a clothing sampler (CS) based on active learning. The main idea is actively selecting valuable clothing images, which can ensure both "clothing diversity with the same identity" and "identity diversity with the similar clothing" in batches, forcing the model to learn features that are independent of clothing. In addition, a multi-clothing loss (MC) is also designed to guide the network to learn clothing-independent features. Experiment results on two cloth-changing datasets show the effectiveness of our proposed CS.