POANET: Parts-Based Occlusion Aware Network for Person Re-Identification

Pratyay Dutta, Padmaja Jonnalagedda, Bir Bhanu

Abstract: Occlusion remains a major challenge in person re-identification (ReID), as it corrupts feature representations and reduces retrieval accuracy. Most existing methods rely on proxy strategies such as random erasing or pose cues, but they lack explicit modeling of where and how occlusion occurs. To address this gap, we propose two key innovations. First, an Occlusion Detection Network (ODN) explicitly estimates the extent and location of occlusion, enabling the ReID backbone to downweight corrupted regions and emphasize features from visible regions. Second, a Part Aware Feature Extractor (PAFE) generates anatomically correct part embeddings (head, torso, legs), allowing the model to dynamically shift focus from occluded to unoccluded parts in a self-supervised manner while preserving global context. The paper also introduces SHHQ-occ, a large-scale dataset of 1.2M images with precise, part-level occlusion annotations, providing a robust foundation for occlusion modeling. Experiments on occluded benchmark datasets: Occluded-ReID, Occluded-Duke, P-DukeMTMC-ReID, PETHZ and holistic reid datasets: Market-1501, CUHK03, DukeMTMC, MSMT17 demonstrate that our approach achieves state-of-the-art accuracy with strong occlusion robustness and efficiency.

External IDs:doi:10.1109/tbiom.2025.3641048