Research map / Continual learning / Domain and prompt-based
Distribution-aware Knowledge Prototyping for Non-exemplar 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 Distribution-aware Knowledge Prototyping (DKP), a non-exemplar lifelong person re-ID method that models per-instance Gaussian distributions, builds identity-level distribution prototypes, and transfers them via a prototype-based KL loss.
Problem
Catastrophic forgetting in lifelong person re-identification when learning from non-stationary data without storing exemplar images (privacy constraints), where existing prototype methods use only feature centers or statistics that ignore instance-level diversity.
Approach
Model the instance-level Gaussian distribution (center and variance) of each image, aggregate these into identity-level distribution prototypes (mean and variance), and transfer them to the next step via a KL-based prototype knowledge transfer loss, followed by model fusion of old and new models.
Key results
- On Training Order-1, DKP reaches Seen-Avg mAP 51.8 / R@1 64.1 versus 43.7/53.7-ish for PatchKD (second best), an 8.1% mAP improvement as claimed by the authors.
- Unseen-Avg mAP/R@1 of 59.2/51.6 under Order-1, outperforming PRAKA and PatchKD on generalization.
- Ablation: adding IDM and DPG&PKT to the baseline raises Seen-Avg mAP from 44.0 to 51.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace PatchKD with DKP | LReID model · Training Order-1 Seen-Avg | mAP: 43.7 -> 51.8 | Helped |
| Replace PatchKD with DKP | LReID model · Training Order-1 Unseen-Avg | mAP: 49.1 -> 59.2 | Helped |
| Replace PatchKD with DKP | LReID model · Training Order-2 Seen-Avg | mAP: 43.2 -> 51.3 | Helped |
| Replace PatchKD with DKP | LReID model · Training Order-2 Unseen-Avg | mAP: 48.6 -> 59.0 | Helped |
| Add IDM (ablation) | Baseline · Seen-Avg | mAP: 44.0 -> 46.1 | Helped |
| Add IDM (ablation) | Baseline · Unseen-Avg | mAP: 46.8 -> 51.6 | Helped |
| Add DPG&PKT (ablation) | Baseline · Seen-Avg | mAP: 44.0 -> 50.0 | Helped |
| Add DPG&PKT (ablation) | Baseline · Unseen-Avg | mAP: 46.8 -> 57.7 | Helped |
| Add IDM and DPG&PKT together (ablation) | Baseline · Seen-Avg | mAP: 44.0 -> 51.8 | Helped |
| Add IDM and DPG&PKT together (ablation) | Baseline · Unseen-Avg | mAP: 46.8 -> 59.2 | Helped |
| Increase the number of stored prototype stages K (ablation) | DKP · LReID | performance | Hurt |
| Increase sampled feature candidate number m (ablation) | DKP · LReID | performance | Helped |
Surprising findings
- PatchKD beats DKP on the first dataset because its strict distillation preserves stability, while DKP's plasticity lags early in training.
- Storing prototypes from more than one previous step degrades performance, so only the immediately previous step is used.
Limitations
- Early-step performance lags behind distillation-based methods such as PatchKD due to slower convergence.
- Prototypes from earlier steps do not help knowledge transfer because they become outdated.
- Final-step performance falls short of PRAKA on CUHK03.
Experimental setup
Datasets: Market1501 CUHK-SYSU DukeMTMC-reID MSMT17-V2 CUHK03 CUHK01 CUHK02 VIPeR PRID i-LIDS GRID SenseReID
Benchmarks: Market1501 CUHK-SYSU DukeMTMC-reID MSMT17-V2 CUHK03 CUHK01 CUHK02 VIPeR PRID i-LIDS GRID SenseReID
Metrics: mAP Rank-1 accuracy average mAP/R@1 over seen and unseen domains
Baselines: LwF SPD PRAKA PRD CRL AKA PatchKD MEGE Joint-Train Finetune
Builds on: PASS PatchKD AKA PRAKA
Closest papers by meaning
- LSTKC: Long Short-Term Knowledge Consolidation for Lifelong Person Re-identification (AAAI 2024 · similarity 0.92)
- DASK: Distribution Rehearsing via Adaptive Style Kernel Learning for Exemplar-Free Lifelong Person Re-Identification (AAAI 2025 · similarity 0.91)
- Lifelong Person Re-Identification via Adaptive Knowledge Accumulation (CVPR 2021 · similarity 0.91)
- DKC: Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification (CVPR 2025 · similarity 0.89)
- Privacy-Aware Lifelong Learning (ICLR 2025 · similarity 0.84)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.84)
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.