CAM2Former: Fusion of Camera-specific Class Activation Map matters for occluded person re-identification

Abstract

Occluded person re-identification (ReID) is challenging since persons are frequently perturbed by various occlusions. Existing mainstream schemes prioritize the alignment of fine-grained body parts by error-prone computation-intensive information, which might come with high estimation error and much computation. To this end, we present the CAMemra-specific Class Activation Map (${\text{CAM}}^2$), designed to identify critical foreground components with interpretability and computational efficiency. Expanding on this foundation, we launched the ${\text{CAM}}^2$-guided Vision Transformer, which is termed ${\text{CAM}}^2$Former, with three core designs. First, we develop Fusion of CAMmera-specific Class Activation Map, termed ${\text{CAM}}^2$Fusion, which consists of positive and negative ${\text{CAM}}^2$ that operate in synergy to capture visual patterns representative of the discriminative foreground components. Second, to enhance the representation ability of pivotal foreground components, we introduce a ${\text{CAM}}^2$Fusion-attention mechanism. This strategy imposes sparse attention weights on identity-agnostic interference discerned by positive and negative ${\text{CAM}}^2$. Third, since the enhancement of foreground representations in ${\text{CAM}}^2$Former depends on camera-specific classifiers, which are not available during inference, we introduce a consistent learning scheme. This design ensures that representations derived from vanilla ViT align consistently with those obtained via ${\text{CAM}}^2$Former. This facilitates the extraction of discriminative foreground representations, circumventing ${\text{CAM}}^2$ dependencies during inference without additional complexity. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art performance on two occluded datasets (Occluded-Duke and Occluded-REID) and two holistic datasets (Market1501 and MSMT17), achieving an R1 of 74.4% and a mAP of 64.8% on Occluded-Dukes.