Research map / Continual learning / Domain and prompt-based
LSTKC: Long Short-Term Knowledge Consolidation for 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 LSTKC, an exemplar-free lifelong person re-ID method that rectifies erroneous old-model relation matrices for short-term distillation and fuses old and new models with knowledge-guided weights for long-term consolidation.
Problem
Catastrophic forgetting in lifelong person re-identification when training sequentially on diverse domains without storing exemplars; standard distillation transfers erroneous old-model knowledge.
Approach
Filters and rectifies false positive/negative entries in the old model's pairwise relation matrix using identity labels before KL-based distillation, then adaptively fuses the parameters of old and new models per step using the relation-matrix difference as the fusion weight.
Key results
- Under training order-1, LSTKC reaches 50.0 seen-avg mAP and 57.0 unseen-avg mAP, improving over PatchKD by 6.3% and 7.9% mAP respectively.
- Under training order-2, LSTKC improves seen-avg mAP by 6.4% and unseen-avg mAP by 9.0% over PatchKD.
- Ablation: R-STKT and E-LTKC each improve over baseline by about 4.8-6.2% mAP alone, and by 8.5-9.8% together.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add R-STKT relation rectification and transfer (ablation) | LSTKC · LReID benchmark, training order-1 | mAP: 40.4 -> 46.1 | Helped |
| add E-LTKC model fusion (ablation) | LSTKC · LReID benchmark, training order-1 | mAP: 40.4 -> 46.3 | Helped |
| add both R-STKT and E-LTKC (ablation) | LSTKC · LReID benchmark, training order-1 | mAP: 40.4 -> 50.0 | Helped |
| replace proposed relation rectification with Max-Min (ablation) | R-STKT · Seen-Avg, training order-1 | mAP: 46.1 -> 43.4 | Hurt |
| replace proposed relation rectification with 1-Minuscule (ablation) | R-STKT · Seen-Avg, training order-1 | mAP: 46.1 -> 25.3 | Hurt |
| replace KL with MSE affinity loss (ablation) | R-STKT · Seen-Avg, training order-1 | mAP: 46.1 -> 40.9 | Hurt |
| replace KL with MAE affinity loss (ablation) | R-STKT · Seen-Avg, training order-1 | mAP: 46.1 -> 41.5 | Hurt |
| replace KL with JS affinity loss (ablation) | R-STKT · Seen-Avg, training order-1 | mAP: 46.1 -> 45.3 | Hurt |
| replace knowledge-guided fusion with fixed 0.5 fusion (ablation) | E-LTKC · Seen-Avg, training order-1 | mAP: 50.0 -> 49.5 | Hurt |
| replace knowledge-guided fusion with time-increasing fusion (ablation) | E-LTKC · Seen-Avg, training order-1 | mAP: 50.0 -> 46.3 | Hurt |
| replace knowledge-guided fusion with time-descending fusion (ablation) | E-LTKC · Seen-Avg, training order-1 | mAP: 50.0 -> 48.9 | Hurt |
| use LSTKC instead of PatchKD | ResNet-50 LReID · Seen-Avg, training order-1 | mAP: 43.7 -> 50.0 | Helped |
Surprising findings
- On the first dataset of each training order, LSTKC is inferior to SOTA methods, which the authors attribute to strict distillation constraints in those methods.
- The method outperforms JointTrain on unseen domains (57.0 vs 49.8 unseen-avg mAP under order-1).
Limitations
- Performs worse than SOTA on the first training dataset due to lack of strict distillation constraints.
- Relies on identity annotations of new data to filter erroneous relations, so it depends on labeled data at each step.
Experimental setup
Datasets: Market-1501 CUHK-SYSU DukeMTMC-reID MSMT17-V2 CUHK03
Benchmarks: Market-1501 CUHK-SYSU DukeMTMC-reID MSMT17-V2 CUHK03 CUHK01 CUHK02 VIPeR PRID i-LIDS GRID SenseReID
Metrics: mAP R@1
Baselines: Finetune JointTrain SPD LwF CRL AKA PatchKD
Builds on: AKA PatchKD LwF
Closest papers by meaning
- Lifelong Person Re-identification via Knowledge Refreshing and Consolidation (AAAI 2023 · similarity 0.94)
- Distribution-aware Knowledge Prototyping for Non-exemplar Lifelong Person Re-identification (CVPR 2024 · similarity 0.92)
- Lifelong Person Re-identification by Pseudo Task Knowledge Preservation (AAAI 2022 · similarity 0.92)
- DKC: Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification (CVPR 2025 · similarity 0.91)
- DASK: Distribution Rehearsing via Adaptive Style Kernel Learning for Exemplar-Free Lifelong Person Re-Identification (AAAI 2025 · similarity 0.89)
- 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.