| Summary: | Occluded person re-identification (ReID) is a person search task aimed at matching occluded pedestrian images with their corresponding full-body images. Local feature-based methods have been demonstrated to be effective in addressing occlusion, as they provide fine-grained information and are especially suitable for representing partially visible body parts. However, training local feature-based models poses two challenges: body part appearance lacks discriminability compared to global appearance and body part misalignment. To address these challenges, this paper proposes LGFNet ( local-global feature Net), a network model that combines global-local feature joint learning with spatial attention maps generated based on self-attention mechanisms. LGFNet allows for automatic adjustment of receptive field sizes. It achieves state-of-the-art results on three public datasets: Market-1501, DukeMTMC reID, and Occluded-Duke. Specifically, LGFNet achieved Rank-1 accuracy of 95. 2% and mAP of 86. 5% on Market-1501, Rank-1 of 90. 2% and mAP of 80. 3% on DukeMTMC-reID, and Rank-1 of 71. 5% and mAP of 60. 9% on Occluded-Duke. These results demonstrate that LGFNet significantly improves accuracy compared to similar methods.
|
|---|
