Research map / Label-efficient and robust learning / Recognition and long-tail
A Simple and Effective Framework for Pairwise Deep Metric Learning
Qi Qi, Yan Yan, Zixuan Wu, Xiaoyu Wang, Tianbao Yang
Read the paperCheck 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 |
|---|---|---|---|
| use DRO-KL_M sampling instead of MS loss | pairwise DML with BN-Inception · Cars-196 | Recall@1: 84.1 -> 86.4 (+2.3) | Improved |
| use DRO-TopK_B instead of MS loss | pairwise DML · Cub-200-2011 | Recall@1: 65.7 -> 68.1 (+2.4) | Improved |
| use DRO-TopK-PN_M instead of MS loss | pairwise DML · In-Shop | Recall@1: 89.7 -> 91.3 (+1.6) | Improved |
| use DRO-KL_M instead of best ensemble ABE | pairwise DML · Cars-196 | Recall@1: 85.2 -> 86.4 (+1.2) | Improved |
| lower gamma in grouped DRO-KL (generalized LS loss) from 1 to 0.001 (ablation) | DRO-KL-G · Cars-196 (d=64) | Recall@1: 70.5 -> 75.0 | Improved |
| lower gamma in grouped DRO-KL from 1 to 0.001 (ablation) | DRO-KL-G · Cub-200-2011 (d=64) | Recall@1: 56.4 -> 56.7 | No clear effect |
| lower gamma in grouped DRO-KL from 1 to 0.01 (ablation) | DRO-KL-G · In-Shop (d=64) | Recall@1: 84.8 -> 85.8 | Improved |
| DRO-KL-G gamma=1 versus LS loss (ablation) | DRO-KL-G · Cars-196 (d=64) | Recall@1: 69.7 -> 70.5 | No clear effect |
| generalized LS loss (gamma=0.001) instead of MS loss (ablation) | DRO-KL-G · Cars-196 (d=64) | Recall@1: 73.2 -> 75.0 | Improved |
| use DRO sampling instead of semi-hard or distance weighted sampling when positive:negative ratio falls (ablation) | DRO variants · imbalance sweep, batch size 80 to 640 | Recall@1 | Improved |
| change embedding dimension from 1024 to 512 (ablation) | DRO-TopK_M · In-Shop (PN ratio 0.006) | Recall@1: 0.9046 -> 0.9058 | No clear effect |
Closest papers by meaning
- Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning (CVPR 2019 · similarity 0.89)
- Improving Deep Metric Learning by Divide and Conquer (TPAMI 2021 · similarity 0.88)
- A unifying mutual information view of metric learning: cross-entropy vs. pairwise losses (ECCV 2020 · similarity 0.88)
- Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (ICML 2020 · similarity 0.88)
- Deep Disentangled Metric Learning (AAAI 2025 · similarity 0.87)
- Potential Field Based Deep Metric Learning (CVPR 2025 · 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.
