Research map / Label-efficient and robust learning / Recognition and long-tail
Person De-reidentification: A Variation-guided Identity Shift Modeling
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.
Introduces person de-reidentification (De-ReID), which unlearns specific persons from a ReID model while preserving matching on accessible persons, using variation-guided identity shift and relation regularization with weakly supervised fine-tuning.
Problem
Person re-identification models can track individuals who have privacy concerns; the task is to make a trained ReID model unable to match specified persons across cameras while keeping its performance on other persons, without relying on labels for accessible persons.
Approach
Instead of unlearning via identity labels or classifier decision boundaries, it fits image-specific style variations for unlearned persons (pushing an image away from its augmented view) while pulling accessible persons' images toward their augmentations, and adds a triplet-style relation constraint and a relation consistency regularizer on unlabeled accessible images.
Key results
- On MSMT17 with MT=100, the method reaches H-Mean 66.7% (R-1T 13.1%, R-1O 67.1%), outperforming Boundary Shrink by 30.4% H-Mean and LIRF by 12.9%.
- On Market-1501 with MT=25, R-1T drops to 10.7% with R-1O 91.1% and H-Mean 84.4% (87.4% with re-ranking).
- Applied to DCFormer, PAT, and PHA, the method gives the best H-Mean among compared De-ReID methods on each backbone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use variation-guided identity shift and relation regularization | ViT-B ReID model · MSMT17, MT=25 | H-Mean | Helped |
| Use variation-guided identity shift and relation regularization | ViT-B ReID model · Market-1501, MT=25 | H-Mean | Helped |
| Add the triplet relation constraint LT_RC (ablation) | Variation-guided identity shift · MSMT17, MT=100 | H-Mean: 63.2 -> 66.7 | Helped |
| Add self-augmented discrimination (ablation) | Variation-guided identity shift · Market-1501, MT=25 | H-Mean: 75.8 -> 84.4 | Helped |
| Add accessible-person variation loss Lo_VIS (ablation) | Variation-guided identity shift · MSMT17, MT=100 | H-Mean: 62.3 -> 66.7 | Helped |
| Add variation-guided identity shift LVIS (ablation) | De-ReID model · MSMT17, MT=100 | Rank-1 accuracy (unlearned persons): 41.3% ... 13.1% | Helped |
| Use more unlabeled accessible-person data | Ours · MSMT17, MT=100 | H-Mean | Helped |
| Add LUPerson auxiliary unlabeled images | Ours · MSMT17, MT=100 | H-Mean: +4.3% | Helped |
Surprising findings
- Existing unlearning methods often get worse unlearning results after re-ranking, e.g., SCRUB's R-1T on unlearned persons rises from 43.8% to 52.3% at MT=25 on MSMT17.
- Adding stronger augmentation alone (LIRF*) does not consistently improve performance, showing that the augmentation alone is not the source of gains.
Limitations
- Performance on accessible persons decreases as the number of unlabeled accessible persons shrinks.
- Relies on a pretrained ReID model and labeled unlearned-person images; the fine-tuning subset is simulated from benchmark splits.
- Evaluated only on a new protocol derived from Market-1501 and MSMT17 (main text), with other datasets deferred to the appendix.
Experimental setup
Datasets: Market-1501 MSMT17 Occ-Duke SYSU-MM01 LUPerson
Benchmarks: Market-1501 MSMT17
Metrics: Rank-1 accuracy on unlearned persons (R-1T) Rank-1 accuracy on accessible persons (R-1O) H-Mean CMC mAP
Baselines: LabelAug Boundary Shrink SCRUB LIRF GS-LoRA
Builds on: GS-LoRA LIRF SCRUB Boundary Shrink RandAugment
Closest papers by meaning
- SANER: Switchable Adapter with Non-parametric Enhanced Routing for Person De-Reidentification (CVPR 2026 · similarity 0.90)
- Single Camera Training for Person Re-Identification (AAAI 2020 · similarity 0.89)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.89)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.89)
- From Laboratory to Real World: A New Benchmark Towards Privacy-Preserved Visible-Infrared Person Re-Identification (CVPR 2025 · similarity 0.87)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 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.