Research map / Label-efficient and robust learning / Recognition and long-tail
Occluded Person Re-identification via Saliency-Guided Patch Transfer
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 Saliency-Guided Patch Transfer (SPT), an online ViT-based data augmentation that transfers salient person patches onto real occluder backgrounds, guided by an occlusion-aware IoU with mask rolling and class-ignoring loss, to improve occluded person re-ID.
Problem
Person re-identification degrades substantially under real-world occlusion, and existing data-driven augmentation (random erasing with random noise) does not mimic realistic occlusion in content or location.
Approach
Uses a learned salient patch selection module to split patches into identity and occlusion sets from real training images, then transfers identity patches onto occluder patches from other images, choosing pairs with an Occlusion-Aware IoU plus mask rolling and ignoring the candidate's identity label in the loss.
Key results
- On Occluded-Duke, ViT-SPT reaches 68.6% Rank-1 and 57.4% mAP vs. 60.7% and 53.0% for the ViT baseline.
- DPM-SPT improves DPM on Occluded-Duke from 71.4% to 74.7% Rank-1 and from 61.8% to 63.0% mAP.
- On holistic Market-1501 and DukeMTMC-reID, ViT-SPT stays roughly on par with the ViT baseline (e.g., 94.5 vs 94.3 Rank-1 on Market-1501).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SPT to DPM | DPM · Occluded-Duke | Rank-1: 71.4 -> 74.7 | Helped |
| Add SPT to DPM | DPM · Occluded-REID | mAP: 79.7 -> 81.1 | Helped |
| Add SPT to ViT | ViT · Occluded-Duke | mAP: 53.0 -> 57.4 | Helped |
| Add SPT to ViT | ViT · Occluded-REID | mAP: 76.7 -> 81.3 | Helped |
| Add SPT to ViT | ViT · Market-1501 | mAP: 86.8 -> 86.2 | Hurt |
| Add SPT to DPM | DPM · DukeMTMC-reID | mAP: 82.6 -> 82.4 | Hurt |
| Add OIoU to random SPT pairing (ablation) | ViT-Base with SPT · Occluded-Duke | Rank-1: 65.3 -> 66.6 | Helped |
| Add mask rolling to OIoU-guided SPT (ablation) | ViT-Base with SPT · Occluded-Duke | Rank-1: 66.6 -> 68.6 | Helped |
| Set SPT probability p to 0.1 instead of 0.2 (ablation) | ViT-SPT · Occluded-Duke | Rank-1: 68.6 -> 67.1 | Hurt |
| Set SPT probability p to 0.3 instead of 0.2 (ablation) | ViT-SPT · Occluded-Duke | Rank-1: 68.6 -> 67.3 | Hurt |
| Set patch budget β to 0.2 instead of 0.3 (ablation) | ViT-SPT · Occluded-Duke | Rank-1: 68.6 -> 65.0 | Hurt |
| Set patch budget β to 0.4 instead of 0.3 (ablation) | ViT-SPT · Occluded-Duke | Rank-1: 68.6 -> 64.1 | Hurt |
Surprising findings
- Even random combination of identity and occluder patches gives a large gain (60.7 to 65.3 Rank-1), attributed to ViT patchify compressing image size.
- Ignoring the candidate's class label in the softmax and triplet losses improves Rank-1 by 3.4 points over standard training.
Limitations
- Gains on holistic benchmarks are negligible or slightly negative (e.g., ViT-SPT 86.2 mAP vs 86.8 on Market-1501).
- Relies on occluders present in the training set, so generated occlusions are limited to those scenes and cameras.
- Hyperparameters (p, beta, alpha1, alpha2) are tuned empirically and sensitive.
Experimental setup
Datasets: Occluded-Duke Occluded-REID Market-1501 DukeMTMC-reID
Benchmarks: Occluded-Duke Occluded-REID Market-1501 DukeMTMC-reID
Metrics: Rank-1 mAP CMC Rank-5 Rank-10
Baselines: Random erasing-style augmentation FED PAT ViT DPM HOReID OAMN Part-Label ISP TransReID
Builds on: ViT (TransReID-style baseline) AdaViT dynamic patch selection random erasing DPM
Closest papers by meaning
- COPE: Consistent Occlusion and Prompt Enhancement Network for Occluded Person Re-identification (CVPR 2026 · similarity 0.88)
- Attention Disturbance and Dual-Path Constraint Network for Occluded Person Re-identification (AAAI 2024 · similarity 0.87)
- Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification (AAAI 2024 · similarity 0.87)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.84)
- Text-Based Occluded Person Re-identification via Multi-Granularity Contrastive Consistency Learning (AAAI 2024 · similarity 0.83)
- Large Occluded Human Image Completion via Image-Prior Cooperating (AAAI 2024 · similarity 0.83)
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.