ABC-Learning: Attention-Boosted Contrastive Learning for unsupervised person re-identification

Unsupervised person re-identification (Re-ID) is a challenging task due to the complexity of the camera view and the variability introduced by pedestrian pose. The irrelevant samples in sample-based contrastive learning pose a potential risk of overfitting which reduces the ability to identify similar features. Therefore, the feature representation relies on the local feature to perceive global content and overfits negative samples. This paper proposes an innovative approach: Attention-Boosted Contrastive Learning (ABC-Learning), which increases region awareness and regularizes in-memory samples. In ABC-Learning, the adjacent content aware (ACA) module is designed to enhance regional feature recognition. It enables local regions of feature maps to re-balance their output by perceiving the informative features of neighboring regions. Additionally, cluster-oriented self-attention (COS) is proposed to regularize the influence of different clusters on the contrastive learning process. The structure of COS reduces the influence of irrelevant clusters through memory data distribution. Finally, a new distance-weighted (DW) memory updating strategy is deployed to keep memory and encoder coherent. DW ensures updating consistency and knowledge transfer through relative distance detection. Our results evaluated on person Re-ID benchmarks demonstrate the promising ability of ABC-Learning to capture the regional feature information and sample weights adjustment.