Research map / Continual learning / Domain and prompt-based
DKC: Differentiated Knowledge Consolidation for Cloth-Hybrid 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 DKC, a non-exemplar framework for cloth-hybrid lifelong person re-ID that adaptively transfers cloth-relevant and cloth-irrelevant knowledge via fine-grained clustering, latent feature-space reconstruction, and dual-level distribution alignment.
Problem
Lifelong person re-identification must match the same person across sequentially arriving cloth-consistent and cloth-changing data, where clothing-based cues become unreliable and cause catastrophic knowledge conflict and forgetting, and existing LReID methods do not balance cloth-relevant and cloth
Approach
Adaptively balance differentiated knowledge by building a pairwise-similarity fusion of old and new models, where the fusion coefficient depends on whether same-identity pairs share fine-grained (DBSCAN-clustered) clothing labels; reconstruct old latent feature space to reduce domain shift; and align instance-level and fine-grained-level distributions to resolve conflicts.
Key results
- On CH-LReID Order-1, DKC reaches 44.2% mAP / 55.8% R1 average, versus 43.8%/55.3% for Joint training and 41.7%/52.8% for DKP.
- On Order-2 DKC reaches 43.7% mAP / 54.6% R1; average forgetting on Order-1 is 3.9% mAP / 4.3% R1, the lowest among compared methods.
- On traditional LReID Order-3 and Order-4, DKC achieves 51.6% mAP average, matching DKP's 51.6% and exceeding it on R1 (63.8% vs 63.7%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DKT (ablation) | Baseline LReID model · CH-LReID Order-1 average | mAP: 41.7 -> 43.1 | Helped |
| Add DKT (ablation) | Baseline LReID model · CH-LReID Order-1 average | R1: 52.8 -> 54.9 | Helped |
| Add LKC (ablation) | Baseline LReID model · CH-LReID Order-1 average | mAP: 41.7 -> 42.6 | Helped |
| Add LKC (ablation) | Baseline LReID model · CH-LReID Order-1 average | R1: 52.8 -> 53.8 | Helped |
| Add DKT and LKC (ablation) | Baseline LReID model · CH-LReID Order-1 average | mAP: 41.7 -> 43.7 | Helped |
| Add DKT and LKC (ablation) | Baseline LReID model · CH-LReID Order-1 average | R1: 52.8 -> 55.0 | Helped |
| Add DDA to Base + DKT + LKC (ablation) | CH-LReID model · CH-LReID Order-1 average | mAP: 43.7 -> 44.2 | Helped |
| Add DDA to Base + DKT + LKC (ablation) | CH-LReID model · CH-LReID Order-1 average | R1: 55.0 -> 55.8 | Helped |
| Use DKC instead of DKP | LReID model · CH-LReID Order-1 average | mAP: 41.7 -> 44.2 | Helped |
| Use DKC instead of DKP | LReID model · CH-LReID Order-2 average | mAP: 40.8 -> 43.7 | Helped |
| Use DKC instead of DKP | LReID model · LReID Order-3 | mAP: 51.8 -> 51.6 | No clear effect |
| Use DKC instead of DKP | LReID model · LReID Order-4 | mAP: 51.3 -> 51.6 | Helped |
Surprising findings
- Sequential fine-tuning with DSIFLF performs worse than plain SFT on several cloth-changing datasets, indicating the cloth-changing CC-ReID method does not transfer to lifelong settings.
Limitations
- Gains over the strongest prior baseline (DKP) on traditional LReID are marginal (51.6 vs 51.6 mAP).
- Relies on non-exemplar distillation and does not use replay; hyperparameters (alpha, beta, mu) tuned on Order-1 ablations.
- Fine-grained labels depend on DBSCAN clustering of features, which may be noisy.
Experimental setup
Datasets: Market-1501 CUHK-SYSU DukeMTMC-ReID MSMT17-V2 CUHK03 LTCC PRCC
Benchmarks: CH-LReID Order-1 CH-LReID Order-2 LReID Order-3 LReID Order-4
Metrics: mAP Rank-1 Average Forgetting
Baselines: LwF AKA PatchKD LSTKC USP DKP DSIFLF CRL
Builds on: DKP PatchKD LSTKC AKA
Closest papers by meaning
- LSTKC: Long Short-Term Knowledge Consolidation for Lifelong Person Re-identification (AAAI 2024 · similarity 0.91)
- Distribution-aware Knowledge Prototyping for Non-exemplar Lifelong Person Re-identification (CVPR 2024 · similarity 0.89)
- DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID (CVPR 2025 · similarity 0.87)
- Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification (CVPR 2025 · similarity 0.87)
- DASK: Distribution Rehearsing via Adaptive Style Kernel Learning for Exemplar-Free Lifelong Person Re-Identification (AAAI 2025 · similarity 0.86)
- Learning Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification (CVPR 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.