Research map / Label-efficient and robust learning / Recognition and long-tail
CA-Jaccard: Camera-aware Jaccard Distance for 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 CA-Jaccard, a camera-aware Jaccard distance that builds camera-aware k-reciprocal neighbors and camera-aware local query expansion to improve pseudo-labeling in unsupervised clustering and re-ranking for person re-ID.
Problem
Jaccard distance, widely used in person re-ID re-ranking and clustering, becomes unreliable under camera variation because intra-camera samples dominate relevant neighbors, introducing negatives and excluding inter-camera positives.
Approach
Find k-reciprocal neighbors separately on intra-camera and inter-camera ranking lists with different k (small intra, large inter), and perform local query expansion by averaging weighted neighbor vectors from both lists so that samples recurring across cameras get higher weight.
Key results
- In clustering with PPLR, CA-Jaccard reaches 86.1% mAP / 94.4% Rank-1 on Market1501, 44.3% / 75.1% on MSMT17, and 45.3% / 90.4% on VeRi-776.
- In re-ranking, BoT+CAJ improves MSMT17 mAP from 50.7 (BoT) to 74.1, versus 66.9 for BoT+KR.
- Applied to CC* in clustering, CA-Jaccard raises MSMT17 mAP from 31.1 to 42.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CA-Jaccard distance for clustering | PPLR · Market1501 | mAP: 84.4 -> 86.1 | Helped |
| Apply CA-Jaccard distance for clustering | PPLR · MSMT17 | mAP: 42.2 -> 44.3 | Helped |
| Apply CA-Jaccard distance for clustering | PPLR · VeRi-776 | mAP: 43.5 -> 45.3 | Helped |
| Use CA-Jaccard for re-ranking | BoT · MSMT17 | mAP: 50.7 -> 74.1 | Helped |
| Use CA-Jaccard for re-ranking instead of ECN | BoT · Market1501 | mAP: 94.4 -> 94.5 | Helped |
| Use CA-Jaccard for re-ranking instead of ECN | BoT · MSMT17 | mAP: 69.0 -> 74.1 | Helped |
| Use CA-Jaccard for re-ranking instead of ECN | BoT · VeRi-776 | mAP: 79.5 -> 81.4 | Helped |
| Replace robust KRNNs with CKRNNs (ablation) | CC* clustering · Market1501 | mAP: 81.0 -> 82.4 | Helped |
| Add CLQE to Jaccard distance (ablation) | CC* clustering · MSMT17 | mAP: 31.1 -> 40.4 | Helped |
| Replace robust KRNNs with CKRNNs (ablation) | BoT+KR re-ranking · MSMT17 | mAP: 66.9 -> 73.1 | Helped |
| Add CLQE to Jaccard distance (ablation) | BoT+KR re-ranking · Market1501 | mAP: 94.2 -> 94.3 | Helped |
| Add CLQE to Jaccard distance (ablation) | BoT+KR re-ranking · MSMT17 | mAP: 66.9 -> 72.0 | Helped |
Surprising findings
- Setting the intra-camera k to 1 (intra-camera neighbors reduced to the sample itself) still achieves high performance, suggesting inter-camera samples are more informative.
Limitations
- Gains depend on the quality of the underlying original distance and pretrained features.
- Hyperparameters (intra/inter k values) were tuned on the evaluated datasets and must be set manually.
Experimental setup
Datasets: Market1501 MSMT17 VeRi-776
Benchmarks: Market1501 MSMT17 VeRi-776
Metrics: mAP CMC Rank-1 Rank-5 Rank-10
Baselines: K-reciprocal re-ranking (KR) ECN CAP CC ICE PPLR RIDE ISE
Builds on: Jaccard distance k-reciprocal re-ranking local query expansion PPLR CC
Closest papers by meaning
- Re-Ranking Person Re-Identification With k-Reciprocal Encoding (CVPR 2017 · similarity 0.89)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.88)
- Semi-Supervised Bayesian Attribute Learning for Person Re-Identification (AAAI 2018 · similarity 0.88)
- A Pose-Sensitive Embedding for Person Re-Identification With Expanded Cross Neighborhood Re-Ranking (CVPR 2018 · similarity 0.87)
- Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification (AAAI 2024 · similarity 0.85)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · 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.