Research map / Continual learning / Domain and prompt-based
Learning Continual Compatible Representation for Re-indexing Free Lifelong 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 C2R, a continual compatible representation method that transfers stored old gallery features into the new model's feature space so lifelong person re-ID works without re-indexing raw images.
Problem
Lifelong person re-identification normally requires re-indexing all historical gallery images after each model update, which is infeasible under data privacy constraints; the paper targets re-indexing-free L-ReID (RFL-ReID) where old gallery features from older models must remain compatible with the
Approach
A Continual Compatible Transfer network maps old gallery features into the new feature space stage by stage, trained with a balanced compatible distillation module (feature alignment plus relation distillation) and a balanced anti-forgetting distillation module (classifier-based and discriminative-consistency losses).
Key results
- On Order-1 under RFL-ReID, C2R reaches average mAP 39.5% and R@1 48.2%, outperforming PatchKD (36.2%/44.4% base-style comparison; reported gain of 3.3/3.8 points over PatchKD*).
- C2R has the lowest average forgetting among compared methods on Order-1 (AF mAP 13.9 vs PatchKD 16.5).
- On the traditional L-ReID task, C2R matches or slightly exceeds PatchKD, with mAP gains of 0.5 (Order-1) and 1.5 (Order-2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace PatchKD* with C2R | RFL-ReID, Order-1 · Order-1 average | mAP: 36.2 -> 39.5 | Helped |
| Replace PatchKD* with C2R | RFL-ReID, Order-1 · Order-1 average | R@1: 44.4 -> 48.2 | Helped |
| Replace PatchKD* with C2R | RFL-ReID, Order-2 · Order-2 average | mAP: 36.1 -> 39.7 | Helped |
| Replace PatchKD* with C2R | RFL-ReID, Order-2 · Order-2 average | R@1: 44.5 -> 48.4 | Helped |
| Replace PatchKD with C2R | L-ReID, Order-1 · Order-1 average | mAP: 43.7 -> 44.2 | Helped |
| Replace PatchKD with C2R | L-ReID, Order-1 · Order-1 average | R@1: 53.7 -> 53.7 | No clear effect |
| Use adaptive rather than fixed am (ablation) | C2R transfer network · Order-1 RFL-ReID | mAP: 39.5 vs 38.2, 38.7, 39.0, 38.4, 38.9, 38.7 | Helped |
| Use adaptive rather than fixed am (ablation) | C2R transfer network · Order-1 RFL-ReID | R@1: 48.2 vs 46.3, 46.9, 47.6, 47.5, 47.8, 47.7 | Helped |
| Add CCT and compatible alignment loss Lca to the baseline (ablation) | C2R · Order-1 RFL-ReID | mAP: 36.2 -> 37.9 | Helped |
| Add CCT and compatible alignment loss Lca to the baseline (ablation) | C2R · Order-1 RFL-ReID | R@1: 44.4 -> 45.6 | Helped |
| Add relationship distillation loss Lcr (ablation) | C2R with Lca · Order-1 RFL-ReID | mAP: 37.9 -> 38.1 | Helped |
| Add relationship distillation loss Lcr (ablation) | C2R with Lca · Order-1 RFL-ReID | R@1: 45.6 -> 46.7 | Helped |
Limitations
- Difficult to capture enough new knowledge for feature transfer when new data is limited (few-shot scenario).
- Noise in new data can hurt transfer; selective transfer using old knowledge is left for future work.
Experimental setup
Datasets: Market-1501 CUHK-SYSU DukeMTMC-ReID MSMT17-V2 CUHK03
Benchmarks: Market-1501 CUHK-SYSU DukeMTMC-ReID MSMT17-V2 CUHK03
Metrics: mAP Rank-1 accuracy average forgetting
Baselines: JointTrain SPD LwF CRL AKA MEGE PatchKD CVS
Builds on: PatchKD compatible training AKA CVS
Closest papers by meaning
- Spatial-Temporal Person Re-Identification (AAAI 2019 · similarity 0.88)
- Lifelong Unsupervised Domain Adaptive Person Re-Identification With Coordinated Anti-Forgetting and Adaptation (CVPR 2022 · similarity 0.88)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.87)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 2024 · similarity 0.86)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.86)
- Multi-Prompts Learning with Cross-Modal Alignment for Attribute-Based Person Re-identification (AAAI 2024 · similarity 0.85)
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.