Research map / Label-efficient and robust learning / Recognition and long-tail
A Metric Learning Reality Check
Kevin Musgrave, Serge Belongie, Ser-Nam Lim
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 |
|---|---|---|---|
| fine-tune with metric learning loss versus pretrained embeddings | BN-Inception · CUB200 | MAP@R (concatenated 512-dim): 14.21 -> 26.53 | Improved |
| use ArcFace instead of classic contrastive loss | BN-Inception · CUB200 | MAP@R (concatenated 512-dim): 26.53 -> 26.45 | No clear effect |
| use MultiSimilarity instead of classic contrastive loss | BN-Inception · Cars196 | MAP@R (concatenated 512-dim): 24.89 -> 28.07 | Improved |
| use MultiSimilarity instead of classic contrastive loss | BN-Inception · CUB200 | MAP@R (concatenated 512-dim): 26.53 -> 24.70 | Hurt |
| use ProxyNCA instead of triplet loss | BN-Inception · SOP | MAP@R (concatenated 512-dim): 43.37 -> 47.22 | Improved |
| add online MS miner to MultiSimilarity loss (ablation) | BN-Inception · CUB200 | MAP@R (concatenated 512-dim): 24.70 -> 26.52 | Improved |
| add online MS miner to MultiSimilarity loss (ablation) | BN-Inception · Cars196 | MAP@R (concatenated 512-dim): 28.07 -> 27.01 | Hurt |
| learn separate margin beta per class instead of a shared beta (ablation) | BN-Inception · CUB200 | MAP@R (concatenated 512-dim): 23.09 -> 23.71 | Improved |
| learn separate margin beta per class instead of a shared beta (ablation) | BN-Inception · Cars196 | MAP@R (concatenated 512-dim): 24.21 -> 23.11 | Hurt |
| concatenate four cross-validation models' 128-dim embeddings to 512-dim instead of averaging separate accuracies (ablation) | Contrastive loss · CUB200 | MAP@R: 21.18 -> 26.53 | Improved |
| evaluate with MAP@R instead of Recall@1 | validation accuracy during training · metric stability | lag-one autocorrelation: 0.73 -> 0.81 | Improved |
| freeze BatchNorm parameters of pretrained model (ablation) | metric learning training · CUB200 | Recall@1: 2 point | Improved |
Closest papers by meaning
- Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (ICML 2020 · similarity 0.84)
- An end-to-end approach for the verification problem: learning the right distance (ICML 2020 · similarity 0.83)
- Mean Field Theory in Deep Metric Learning (ICLR 2024 · similarity 0.82)
- Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.82)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · similarity 0.82)
- Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning (ICLR 2024 · similarity 0.81)
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.
