A Keypoint‐Guided Feature Partition Network for Occluded Person Re‐Identification

Abstract

Existing occluded person re‐identification methods employ hard or soft partition strategies to explore fine‐grained information. However, the hard partition strategy which extracts region‐level features may impair the semantic connectivity of correlated human body parts. A pose‐guided soft partition establishes correlations among human keypoints, while the generated pixel‐level embeddings may lose the surrounding semantic information. In this paper, we propose a keypoint‐guided feature partition (KGFP) method that consists of a feature extractor, a hard partition branch, and a soft partition branch. Specifically, we adopt a vision transformer and a pose estimator to extract features and keypoint information. In the hard partition branch, we partition features into distinct groups and classify them into nonoccluded, semi‐occluded, and occluded features to obtain region‐level features and filter out occlusions. Furthermore, we design a dissimilarity loss to reduce the similarity between semi‐occluded and occluded features. In the soft partition branch, we introduce a graph attention network and consider global and keypoint embeddings as nodes of a graph to discover interrelationships. Additionally, we formulate image alignment as a graph matching problem and propose a feature alignment‐based graph to reduce position misalignment. Extensive experiments demonstrate that the proposed method achieves superior performance compared to state‐of‐the‐art methods on Occluded‐DukeMTMC, Markt1501, and DukeMTMC‐reID.