Research map / Label-efficient and robust learning / Recognition and long-tail
Revisiting Training Strategies and Generalization Performance in Deep Metric Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add ρ-regularization to Margin (D, β=1.2) (ablation) | ResNet50 DML · CUB200-2011 | Recall@1: 63.09±0.46 -> 64.93±0.42 | Helped |
| Add ρ-regularization to MultiSimilarity (ablation) | ResNet50 DML · CARS196 | Recall@1: 81.68±0.19 -> 82.37±0.13 | Helped |
| Add ρ-regularization to Margin (D, β=1.2) (ablation) | ResNet50 DML · SOP | Recall@1: 78.43±0.07 -> 78.52±0.10 | Helped |
| Use ρ-regularization for ranking-based objectives (ablation) | Ranking-based DML models · CARS196 | Recall@1: more then 5% | Helped |
| Increase batch size (ablation) | DML training · CUB200-2011, CARS196, SOP | Aggregated normalized performance | Helped |
| Vary weight decay (ablation) | DML training · CUB200-2011, CARS196, SOP | Aggregated normalized performance: 5% | Mixed |
| Use larger training images (ablation) | DML training · CUB200-2011, CARS196, SOP | Aggregated normalized performance | Helped |
| Use advanced color and brightness augmentation (ablation) | DML training · CUB200-2011, CARS196, SOP | Aggregated normalized performance | Mixed |
| Use SPC-2 or FRD sampling instead of SPC-4 (ablation) | DML mini-batch sampling · CUB200-2011, CARS196, SOP | Aggregated normalized performance: 1.5% | Helped |
| Increase embedding dimensionality toward 512 (ablation) | ResNet50 DML · CUB200-2011, CARS196, SOP | Aggregated normalized performance | Mixed |
| Reduce compression of learned embeddings | DML models · CUB200-2011, CARS196, SOP | Recall@1: ≤ -0.63 | Helped |
| Encourage higher embedding compression (ablation) | DML models · Standard benchmark datasets | Recall performance | Hurt |
Closest papers by meaning
- Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning (NeurIPS 2021 · similarity 0.91)
- Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning (ICLR 2024 · similarity 0.89)
- Towards Improved Proxy-Based Deep Metric Learning via Data-Augmented Domain Adaptation (AAAI 2024 · similarity 0.88)
- Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.88)
- Deep Disentangled Metric Learning (AAAI 2025 · similarity 0.88)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · similarity 0.87)
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.