Research map / Label-efficient and robust learning / Recognition and long-tail
Graph Sampling Based Deep Metric Learning for Generalizable Person Re-Identification
Shengcai Liao, Ling Shao
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace random PK sampling with nearest-neighbor graph sampling | QAConv · Cross-dataset person re-identification, including MSMT17 training → CUHK03-NP | Rank-1: 18.6 -> 20.9 | Improved |
| Replace random PK sampling with nearest-neighbor graph sampling | QAConv · RandPerson training → MSMT17 | Rank-1: 40.3 -> 45.1 | Improved |
| Use graph sampling instead of PK sampling | OSNet-IBN · MSMT17 (all) training → Market-1501 | Rank-1: 67.9 -> 71.3 | Improved |
| Use graph sampling instead of PK sampling | TransMatcher · Market-1501 training → CUHK03-NP | Rank-1: 22.9 -> 22.2 | Hurt |
| Replace hard triplet loss with binary cross-entropy loss (ablation) | QAConv-GS · Market-1501 training → CUHK03-NP | Rank-1: 19.1 -> 16.4 | Hurt |
| Use 256×128 instead of 384×128 input (ablation) | QAConv-GS · Market-1501 training → CUHK03-NP | Rank-1: 19.1 -> 16.9 | Hurt |
| Use class centers for graph construction instead of one sampled example per class (ablation) | QAConv-GS · Market-1501 training → CUHK03-NP | Rank-1: 19.1 -> 15.4 | Hurt |
| Use class centers for graph construction instead of one sampled example per class (ablation) | QAConv-GS · MSMT17 training → CUHK03-NP | Rank-1: 20.9 -> 15.3 | Hurt |
| Apply gradient clipping (ablation) | QAConv-GS · MSMT17 training | generalization performance | No clear effect |
| Apply gradient clipping (ablation) | QAConv-GS · Market-1501 and RandPerson training | generalization performance | Improved |
| Increase triplet-loss margin above 32 (ablation) | QAConv-GS · MSMT17 training | mAcc | Hurt |
| Increase the number of images per class beyond K=2 with hard triplet loss (ablation) | QAConv-GS · Training convergence | convergence | Hurt |
Closest papers by meaning
- Embedding Deep Metric for Person Re-identification: A Study Against Large Variations (ECCV 2016 · similarity 0.90)
- MVP Matching: A Maximum-Value Perfect Matching for Mining Hard Samples, With Application to Person Re-Identification (ICCV 2019 · similarity 0.88)
- Deep Graph Metric Learning for Weakly Supervised Person Re-Identification (TPAMI 2021 · similarity 0.87)
- Hard-Aware Point-to-Set Deep Metric for Person Re-identification (ECCV 2018 · similarity 0.87)
- Efficient Online Local Metric Adaptation via Negative Samples for Person Re-Identification (ICCV 2017 · similarity 0.86)
- Maximum Margin Metric Learning Over Discriminative Nullspace for Person Re-identification (ECCV 2018 · similarity 0.86)
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.
