Research map / Label-efficient and robust learning / Recognition and long-tail
Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FPC, a ViT-based framework that prunes occluder/background tokens via class-token attention, matches gallery neighbors with a cosine+EMD distance, and consolidates the occluded query CLS token with neighbor features via a transformer decoder.
Problem
Occluded person re-identification, where occluders and background contaminate features and the target's essential features may be partially absent, especially under heavy occlusion and with other persons as occluders.
Approach
Replaces human-structure cues (pose, segmentation) with attention-based token sparsification in a ViT sparse encoder, then recovers pruned query features by consolidating multi-view features from k-nearest gallery neighbors matched by a combined image-level cosine and patch-level EMD distance.
Key results
- On Occluded-Duke, FPC reaches 76.7% Rank-1 and 72.8% mAP, outperforming prior methods by at least 6.0% Rank-1 and 8.6% mAP per the paper.
- Feature consolidation (C) adds 7.2% Rank-1 and 11.5% mAP over the sparse-encoder-only variant in the ablation on Occluded-Duke.
- Token sparsification with keep rate 0.8 cuts sparse encoder FLOPs by 25% while improving Rank-1 from 76.0% to 76.7% versus no pruning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sparse encoder (ablation) | FPC baseline architecture · Occluded-Duke | Rank-1: 65.2 -> 68.6; 3.4% improvement | Helped |
| add feature consolidation (ablation) | sparse encoder · Occluded-Duke | mAP: 60.1 -> 71.6; 11.5% | Helped |
| add multi-view feature matching (ablation) | sparse encoder + consolidation · Occluded-Duke | mAP: 71.6 -> 72.8; 1.2% | Helped |
| set keep rate to 0.8 instead of 1.0 (ablation) | FPC sparse encoder · Occluded-Duke | mAP: 72.3 -> 72.8; +0.5 | Helped |
| replace full gallery image tokens with [cls] tokens (ablation) | FPC · Occluded-Duke | mAP: 72.8% to 70.2% | Hurt |
| add token sparsification at keep rate 0.9 | Transreid · Occluded-Duke | mAP: 59.5 -> 60.0 | Helped |
| move sparsification layers from [3,6,9] to [1,4,7] (ablation) | FPC sparse encoder · Occluded-Duke | Rank-1: 76.7 -> 75.3 | Hurt |
| use two-stage cosine filtering and EMD instead of exhaustive EMD (ablation) | FPC matching module · Occluded-Duke | Rank-1: 76.7 -> 76.8 | No clear effect |
Surprising findings
- Random patch dropping still achieves relatively high performance, suggesting robustness of the sparse encoder to information loss.
- Using only the [cls] token for gallery features reduces memory to 0.05G with only a modest drop (72.8% to 70.2% mAP), at a cost that the paper highlights as acceptable.
Limitations
- Patch drop rate is fixed rather than adaptive, as acknowledged by the authors.
- Underperforms PAT and FRT in Rank-1 on Partial-ReID, attributed to ViT overfitting on small datasets and weaker cross-domain generalization.
- Relies on gallery neighbor retrieval over the full training gallery, raising storage and speed concerns for real-world deployment.
Experimental setup
Datasets: Occluded-Duke Occluded-ReID Partial-ReID Market-1501
Benchmarks: Occluded-Duke Occluded-ReID Partial-ReID Market-1501
Metrics: Rank-1 mAP FLOPs
Baselines: DSR PGFA HOReID OAMN PAT TransReID PFD FED RFCnet FRT Yu et al.
Builds on: TransReID EviT token sparsification Feature Recovery Transformer (FRT) Sinkhorn algorithm k-reciprocal re-ranking
Closest papers by meaning
- Feature Erasing and Diffusion Network for Occluded Person Re-Identification (CVPR 2022 · similarity 0.90)
- COPE: Consistent Occlusion and Prompt Enhancement Network for Occluded Person Re-identification (CVPR 2026 · similarity 0.90)
- Adversarially Occluded Samples for Person Re-Identification (CVPR 2018 · similarity 0.90)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.88)
- Text-Based Occluded Person Re-identification via Multi-Granularity Contrastive Consistency Learning (AAAI 2024 · similarity 0.88)
- Occluded Person Re-identification via Saliency-Guided Patch Transfer (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.