Research map / Continual learning / Domain and prompt-based
DASK: Distribution Rehearsing via Adaptive Style Kernel Learning for Exemplar-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 DASK, an exemplar-free lifelong person re-ID method that learns an instance-adaptive kernel predictor (AKPNet) to rehearse old-domain image styles from new data, combined with cross-instance similarity distillation.
Problem
Lifelong person re-identification suffers from catastrophic forgetting across domain shifts; replay methods violate data privacy by storing exemplars, and knowledge distillation gives limited anti-forgetting capacity.
Approach
Model the old domain distributions by learning a Distribution Rehearser (AKPNet) that predicts per-instance convolution kernels to transform arbitrary-style images into the current domain style; the previous step's AKPNet then synthesizes old-style versions of new data for joint consolidation training with the old and new ReID models.
Key results
- On Training Order-1, DASK reaches Seen-Avg 55.4 mAP / 69.3 R@1 and UnSeen-Avg 65.3 mAP / 58.4 R@1, improving over DKP by 3.6%/5.2% on seen-average.
- On Training Order-2, DASK achieves Seen-Avg 65.5 mAP / 70.0 R@1 and UnSeen-Avg 65.5 mAP / 57.9 R@1, with 5.8%/6.8% seen-average gains over DKP.
- Without exemplar replay, DASK exceeds re-implemented replay baselines (e.g., PTKP) by 4.7%/5.7% and 6.4%/7.2% Seen-Avg mAP/R@1 across the two orders.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DASK instead of DKP | LReID benchmark, Order-1 · Seen-Avg | mAP: 3.6% | Helped |
| Use DASK instead of DKP | LReID benchmark, Order-1 · UnSeen-Avg | mAP: 6.1% | Helped |
| Use DASK instead of DKP | LReID benchmark, Order-2 · Seen-Avg | mAP: 5.8% | Helped |
| Use DASK instead of DKP | LReID benchmark, Order-2 · UnSeen-Avg | mAP: 6.5% | Helped |
| Add distribution rehearsal using DASK's adaptive transfer kernel rather than training on new data only (ablation) | LReID, Order-1 · Seen-Avg | mAP: 48.7 -> 55.4 | Helped |
| Use distribution-augmented data instead of rehearsed old-style data (ablation) | ReID model training · Seen-Avg | mAP: 55.4 -> 48.6 | Hurt |
| Use one shared convolution kernel for all images instead of instance-adaptive kernels (ablation) | ReID model training · Seen-Avg | mAP: 55.4 -> 48.6 | Hurt |
| Predict only mean and standard deviation rather than an adaptive transfer kernel (ablation) | ReID model training · Seen-Avg | mAP: 55.4 -> 51.2 | Hurt |
| Add rehearsed ReID loss L_ReID* (ablation) | Baseline LReID training · Seen-Avg (Order-1) | mAP: 48.7 -> 55.0 | Helped |
| Add rehearsed similarity distillation loss L_SKD* (ablation) | Baseline LReID training · Seen-Avg (Order-1) | mAP: 48.7 -> 49.9 | Helped |
| Use both rehearsed losses rather than baseline losses alone (ablation) | Baseline LReID training · UnSeen-Avg (Order-1) | mAP: 55.5 -> 65.3 | Helped |
| Increase the number of style-transfer kernels above one (ablation) | DASK style transfer · Seen and unseen domains | model performance | Hurt |
Surprising findings
- Preserving all historical AKPNet models gives comparable performance to keeping only the single previous one.
- Stacking multiple predicted kernel layers (N_k>1) degrades performance, with N_k=1 best.
Limitations
- Anti-forgetting designs partly limit learning of new data, yielding inferior results to PRAKA on the last CUHK03 subset.
- Multi-layer kernel prediction overfits seen textures and degrades performance.
- Generalization to unseen domains is sensitive to the loss weight beta due to balancing diverse styles.
Experimental setup
Datasets: Market-1501 DukeMTMC-reID CUHK-SYSU MSMT17-V2 CUHK03 CUHK01 CUHK02 VIPeR PRID i-LIDS GRID SenseReID
Benchmarks: LReID benchmark (Training Order-1) LReID benchmark (Training Order-2)
Metrics: mAP Rank@1 Seen-Avg UnSeen-Avg
Baselines: LwF SPD PRD PRAKA FCS CRL AKA PatchKD MEGE ConRFL CKP LSTKC
Builds on: LwF CoP PatchKD DKP
Closest papers by meaning
- Distribution-aware Knowledge Prototyping for Non-exemplar Lifelong Person Re-identification (CVPR 2024 · similarity 0.91)
- Dynamic Magic: Unleashing Restricted Knowledge for Lifelong Person Re-Identification (CVPR 2026 · similarity 0.90)
- LSTKC: Long Short-Term Knowledge Consolidation for Lifelong Person Re-identification (AAAI 2024 · similarity 0.89)
- Lifelong Person Re-Identification via Adaptive Knowledge Accumulation (CVPR 2021 · similarity 0.89)
- DKC: Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification (CVPR 2025 · similarity 0.86)
- Learning Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification (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.