Research map / Label-efficient and robust learning / Recognition and long-tail
Attention Disturbance and Dual-Path Constraint Network 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 ADP, a ViT-based occluded person re-ID method with an attack-style Attention Disturbance Mask (ADM) that generates realistic feature-level occlusions and a Dual-Path Constraint (DPC) module that uses holistic images as extra supervision.
Problem
Attention networks for occluded person re-identification trained with background-based artificial occlusion generalize poorly to realistic occlusions at test time.
Approach
An attack-style learnable occlusion mask (ADM) pasted with background patches is optimized to pull the class token's attention toward the occluded region, simulating realistic obstacles; a parameter-shared dual-path (holistic and occluded) training scheme with asymmetric classifiers and an interaction loss supplies extra supervision. Both modules are used only in training, so inference uses the pl
Key results
- ADP reaches 74.5% Rank-1 and 63.8% mAP on Occluded-Duke with a sliding-window ViT backbone (ADP*), versus 66.4% Rank-1 for TransReID*.
- On Occluded-Duke, ADM alone raises ViT baseline Rank-1 from 59.7% to 66.2%, and adding DPC gives 72.2%.
- On Occluded-REID, ADP* achieves 89.2% Rank-1 and 85.1% mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ADM-based artificial occlusion (ablation) | ViT baseline · Occluded-Duke | Rank-1: 59.7 -> 66.2; +$6.5\%$ | Helped |
| Add ADM-based artificial occlusion (ablation) | ViT baseline · Occluded-Duke | mAP: 49.8 -> 57.7; +$7.9\%$ | Helped |
| Add DPC after ADM (ablation) | ViT baseline + ADM · Occluded-Duke | Rank-1: 66.2 -> 72.2 | Helped |
| Add DPC after ADM (ablation) | ViT baseline + ADM · Occluded-Duke | mAP: 57.7 -> 60.6 | Helped |
| Replace single-path shared angular-softmax structure with asymmetric dual-path structure (ablation) | ADM model · Occluded-Duke | Rank-1: + $1.0\%$ | Helped |
| Replace single-path shared angular-softmax structure with asymmetric dual-path structure (ablation) | ADM model · Occluded-Duke | mAP: + $0.5\%$ | Helped |
| Add interaction loss L_itr (ablation) | ADM + asymmetric dual-path model · Occluded-Duke | Rank-1: + $1.3\%$ | Helped |
| Add global triplet loss L_tri (ablation) | ADM + DPC + L_itr model · Occluded-Duke | Rank-1: + $0.4\%$ | Helped |
| Use ADM instead of random erasing (ablation) | ViT baseline · Occluded-Duke | Rank-1: + $5.1\%$ | Helped |
| Use ADM instead of background occlusion (ablation) | ViT baseline · Occluded-Duke | Rank-1: + $1.4\%$ | Helped |
| Add DPC to random-erasing training (ablation) | ViT baseline with random erasing · Occluded-Duke | Rank-1: 61.1\% to 66.7\% | Helped |
| Add DPC to background-occlusion training (ablation) | ViT baseline with background occlusion · Occluded-Duke | Rank-1: + $5.6\%$ | Helped |
Surprising findings
- Background-patch occlusion alone performs better than random erasing, but the learned ADM still beats it by 1.4% Rank-1 on Occluded-Duke.
Limitations
- Gains on holistic datasets are only comparable to SOTA, not clearly superior (e.g., Market-1501 mAP below SAP and PFD*).
- Evaluation relies on attention visualizations and t-SNE, without a quantitative measure of realism of the generated occlusions.
Experimental setup
Datasets: Occluded-Duke Market-1501 DukeMTMC-reID Occluded-REID ImageNet
Benchmarks: Occluded-Duke Occluded-REID Market-1501 DukeMTMC-reID
Metrics: Rank-1 mAP Rank-5 Rank-10
Baselines: PCB DSR FPR PVPM HOReID OAMN Part-Label ISP PRE-Net CAAO PAT TransReID
Builds on: TransReID random erasing ViT
Closest papers by meaning
- COPE: Consistent Occlusion and Prompt Enhancement Network for Occluded Person Re-identification (CVPR 2026 · similarity 0.89)
- Feature Erasing and Diffusion Network for Occluded Person Re-Identification (CVPR 2022 · similarity 0.88)
- Occluded Person Re-identification via Saliency-Guided Patch Transfer (AAAI 2024 · similarity 0.87)
- VRSTC: Occlusion-Free Video Person Re-Identification (CVPR 2019 · similarity 0.87)
- Pose-Guided Feature Disentangling for Occluded Person Re-identification Based on Transformer (AAAI 2022 · similarity 0.86)
- Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification (AAAI 2024 · similarity 0.86)
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.