Research map / Label-efficient and robust learning / Recognition and long-tail
From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization
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 Pose2ID, a training-free person re-ID framework that centralizes features by aggregating same-identity features from identity-guided pose-conditioned diffusion-generated images and from mutual-nearest neighbors (NFC).
Problem
Person re-identification features are corrupted by noise from background, occlusion, and model limitations; existing ReID methods require training and generated data often loses identity consistency.
Approach
Features of the same identity cluster around an identity center approximately as a normal distribution, so averaging same-identity features (from identity-guided pose-varied generated images and from neighbor features) reduces individual noise while preserving the distribution for re-ranking; the generator is guided by ReID identity features via an IFR module.
Key results
- Applied to an ImageNet pre-trained ViT without ReID training, IPG reaches 52.81 mAP / 78.92 Rank-1 on Market1501, and IPG+NFC reaches 57.27 mAP / 82.39 Rank-1.
- On Market1501 with TransReID (w/o camid), the method raises mAP from 79.88 to 90.39 and Rank-1 from 91.48 to 94.74.
- Claims new SOTA without re-ranking: CLIP-ReID mAP 94.9 / Rank-1 97.3 on Market1501; KPR mAP 89.34 on Occluded-ReID; SAAI 76.44 mAP on SYSU-MM01 all-search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Pose2ID to TransReID (w/o camid) | TransReID · Market1501 | mAP: 79.88 -> 90.39 | Helped |
| add Pose2ID to CLIP-ReID | CLIP-ReID · Market1501 | mAP: 89.8 -> 94.9 | Helped |
| add Pose2ID to KPR | KPR · Occluded-ReID | mAP: 79.05 -> 89.34 | Helped |
| add Pose2ID to SAAI | SAAI · SYSU-MM01 All-search | mAP: 71.81 -> 76.44 | Helped |
| add IPG (ablation) | ImageNet-pretrained model without ReID training · Market1501 | mAP: 3.34 -> 52.81 | Helped |
| add NFC after IPG (ablation) | ImageNet-pretrained model without ReID training · Market1501 | mAP: 52.81 -> 57.27 | Helped |
| add NFC only (ablation) | TransReID · Market1501 | mAP: 79.88 -> 83.92 | Helped |
| add IPG only (ablation) | TransReID · Market1501 | mAP: 79.88 -> 88.02 | Helped |
| apply NFC to gallery and query rather than neither (ablation) | TransReID · Market1501 | mAP: 79.88 -> 83.92 | Helped |
| apply IPG to gallery and query rather than neither (ablation) | TransReID · Market1501 | mAP: 79.88 -> 88.02 | Helped |
| increase generated images from 0 to 8 per gallery sample (ablation) | TransReID with IPG on gallery · Market1501 | mAP: 79.88 -> 84.65 | Helped |
| increase generated images from 0 to 8 per query sample (ablation) | TransReID with IPG on query · Market1501 | mAP: 79.88 -> 82.18 | Helped |
Surprising findings
- An ImageNet pre-trained ViT with no ReID training reaches competitive retrieval (52.81 mAP) after feature centralization.
Limitations
- Relies on generating 8 images per test image, adding inference cost.
- Generator uses pose estimation and data cleaning; performance depends on generated image quality and hyperparameter eta/k choices tuned per dataset.
- Test-time evaluation is done without camera IDs, a setting that differs from standard protocol for some models.
Experimental setup
Datasets: Market1501 MSMT17 Mars SYSU-MM01 Occluded-ReID
Benchmarks: Market1501 SYSU-MM01 Occluded-ReID
Metrics: mAP Rank-1 ID2
Baselines: TransReID CLIP-ReID KPR BPBReID SAAI PMT k-reciprocal re-ranking
Builds on: Stable Diffusion ControlNet TransReID k-reciprocal re-ranking
Closest papers by meaning
- Adversarially Occluded Samples for Person Re-Identification (CVPR 2018 · similarity 0.91)
- Learning Disentangled Representation for Robust Person Re-identification (NeurIPS 2019 · similarity 0.90)
- Person30K: A Dual-Meta Generalization Network for Person Re-Identification (CVPR 2021 · similarity 0.90)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.89)
- Person De-reidentification: A Variation-guided Identity Shift Modeling (CVPR 2025 · similarity 0.89)
- Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification (AAAI 2024 · similarity 0.88)
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.