Camera Proxy Based Contrastive Learning with Hard Sampling for Unsupervised Person Re-identification.

Yimin Liu,Meibin Qi,Qiang Wu,Yanfang Yang,Xiaohong Li,Jian Zhang

DOI: https://doi.org/10.1109/icme55011.2023.00413

2023-01-01

Abstract:Because of the advantages of dealing with large-scale unlabelled data, unsupervised learning has recently attracted more attention for person re-identification. Particularly, the combination of the unsupervised learning paradigm with contrastive learning shows promising efficiency in network optimization. This work adopts the successful camera-aware contrastive learning approach and further explores its capability on the camera proxy level to improve the data pair consistency. Thus, it is more robust to the camera change, which still challenges the unsupervised person re-identification. This work proposed a Camera Proxy-based Contrastive Learning framework, which explicitly considers inter-camera scenario and intra-camera scenario. Moreover, this work is motivated by the strategy of selecting a hard negative sample in triplet loss learning and further extends it to contrastive learning for both negative and positive pair creation on the camera proxy level. Extensive experiments demonstrate the superiority of the proposed framework over state-of-the-art approaches on purely unsupervised re-identification.