Pose-guided counterfactual inference for occluded person re-identification

Occluded person re-identification (ReID) is a challenging task as the body is partially occluded by obstacles in crowd scenarios. Many previous studies employed an attention mechanism to focus on fine-grained local information with conventional likelihood while ignoring the inherent causality between the final prediction results and attention, especially occluded person always possesses biased clues. To address this problem, we propose a Pose-Guided Multi-Attention Network (PGMA-Net) for occluded person ReID in an end-to-end manner. PGMA-Net contains two main novel components: Pose-Guided Counterfactual Inference Branch (PGCIB) and Striped- and Patched-Attention Module (SPAM). The PGCIB jointly explores the causality between the predicted identities and input clues to alleviate the negative effects brought by occluded bias. Specifically, the counterfactual inference can directly guide the attention learning process via the counterfactual intervention. The SPAM generates a set of attention vectors for storing part prototypes over multiple rounds of attention. We empirically demonstrate that PGMA-Net can improve the recognition in both occluded and non-occluded ReID. With the above designs, our framework achieves 52.2% in mAP and 62.0% in top-1 on the Occluded-DukeMTMC dataset, surpassing the baseline by a large margin.