Reproduce “Dual-Prompt CLIP with Hybrid Visual Encoders for Occluded Person Re-Identification” on OpenResearch

About this paper

Zhangjian Ji, Shaotong Qiao, Kai Feng, Wei Wei

The DPL-ReID framework improves occluded person re-identification by introducing realistic occlusion augmentation, dual prompt learning for combining holistic and occlusion-aware semantics, and a hybrid visual encoder. It achieved mAP gains of up to 9.5% and Rank-1 accuracy gains of up to 13.9% over previous methods on occluded datasets like Occluded-Duke with a ViT-B/16 backbone.

Key insights

• Generating realistic and semantically consistent occlusion augmentations, rather than simplistic ones, is crucial for effectively bridging the domain gap in occluded ReID.
• Explicitly incorporating occluder semantic information through dual prompt learning allows vision-language models to derive richer contextual understanding and maintain robustness in occluded scenarios.
• Fusing global features from a lightweight network with the CLIP visual encoder helps mitigate its reliance on localized cues, leading to more comprehensive pedestrian representations useful for both occluded and holistic ReID.

Summary by alphaXiv.