Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Group Re-identification via Adaptive Clustering-Driven Progressive Learning
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 ACPL, an unsupervised group re-identification method that combines a quasi-distance-based group adaptive clustering (GAC) with a global dynamic prototype update (GDPU) to produce cleaner pseudo-labels and mine hard samples.
Problem
Unsupervised group re-identification, where member variations within groups make clustering-based pseudo-labels unstable and noisy, without any cross-camera identity labels.
Approach
Introduces a quasi-distance vector between groups (Hausdorff-inspired, combining person-level min-distances and the mean-feature distance), uses the maximum-margin boundary in sorted distances to form group clusters, and replaces fixed momentum prototype updates with a dynamic weight that emphasizes hard samples far from their prototype.
Key results
- On CSG, ACPL reaches 78.0 mAP and 80.5 R1, outperforming unsupervised and coarse-label baselines but trailing fully supervised 3DT+ (94.4 mAP).
- On i-LIDS MCTS, ACPL achieves 66.1 mAP and 65.3 R1, surpassing the listed fully supervised methods.
- On SYSU-Group, ACPL achieves 85.7 mAP and 96.5 R1, exceeding SVIGR (76.7 mAP) and PCPGU (97.4 R1 is higher for PCPGU, so ACPL is comparable on R1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GAC instead of DBSCAN for group clustering (ablation) | ACPL · CSG, RoadGroup, i-LIDS MCTS | mAP | Helped |
| use overall plus individual features in the quasi-distance vector instead of overall features only (ablation) | ACPL GAC · CSG | mAP: 55.0 -> 78.0 | Helped |
| use overall plus individual features in the quasi-distance vector instead of individual features only (ablation) | ACPL GAC · CSG | mAP: 76.3 -> 78.0 | Helped |
| use overall plus individual features in the quasi-distance vector instead of the most similar person feature distance (ablation) | ACPL GAC · CSG | mAP: 73.4 -> 78.0 | Helped |
| use GDPU instead of momentum update (MU) (ablation) | ACPL person-level training · CSG, RoadGroup, i-LIDS MCTS | mAP and R1: +1.2; +2.1; +2.6 | Helped |
| use GDPU instead of momentum update (MU) (ablation) | ACPL group-level training · CSG, RoadGroup, i-LIDS MCTS | mAP and R1: +2.0; +2.4; +2.5 | Helped |
| use GDPU instead of hardest-sample momentum update (MU-H) (ablation) | ACPL person-level training · CSG, RoadGroup, i-LIDS MCTS | mAP and R1: +1.7; +1.6; +3.4 | Helped |
| use GDPU instead of hardest-sample momentum update (MU-H) (ablation) | ACPL group-level training · CSG, RoadGroup, i-LIDS MCTS | mAP and R1: +2.3; +2.5; +2.6 | Helped |
Surprising findings
- On CSG, the unsupervised method lags well behind supervised methods, which the authors attribute to many distractor pedestrians not belonging to the group.
- Spectral and K-means clustering given the true number of identities are competitive, with GAC still outperforming them.
Limitations
- Performance gap to supervised methods on CSG due to pedestrians outside the group in images.
- Relies on person-level DBSCAN and fixed hyperparameters (eps=0.6, epsilon_c=3, Nc=10).
Experimental setup
Datasets: CSG RoadGroup i-LIDS MCTS SYSU-Group
Benchmarks: CSG RoadGroup i-LIDS MCTS SYSU-Group
Metrics: mAP CMC Rank-1 Rank-5 Rank-10 Fowlkes-Mallows score adjusted rand score adjusted mutual info score V-measure
Baselines: MGR MACG DotSCN SVIGR BDFNet SOT 3DT C-BRO Covariance BSC+CM PREF LIMI
Builds on: Cluster Contrast DBSCAN Hausdorff distance momentum update (MoCo)
Closest papers by meaning
- Group-aware Label Transfer for Domain Adaptive Person Re-identification (CVPR 2021 · similarity 0.89)
- Similarity Metric Learning for RGB-Infrared Group Re-Identification (CVPR 2023 · similarity 0.87)
- Divide-and-Regroup Clustering for Domain Adaptive Person Re-identification (AAAI 2022 · similarity 0.87)
- Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions (ICLR 2025 · similarity 0.85)
- Catalyst for Clustering-Based Unsupervised Object Re-identification: Feature Calibration (AAAI 2024 · similarity 0.85)
- Project-Probe-Aggregate: Efficient Fine-Tuning for Group Robustness (CVPR 2025 · similarity 0.83)
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.