Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking Preventing Class-Collapsing in Metric Learning With Margin-Based Losses
Elad Levi, Tete Xiao, Xiaolong Wang, Trevor Darrell
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 |
|---|---|---|---|
| replace random positive sampling with easy positive sampling (EPS) | Triplet loss · MNIST test digits (unseen) | Recall@1: 35.2 -> 42.3 | Improved |
| add EPS to margin loss with distance sampling | Dist-Margin · Cars-196 | Recall@1: 79.6 -> 83.6 | Improved |
| add EPS to margin loss with distance sampling | Dist-Margin · CUB-200-2011 | Recall@1: 63.6 -> 64.7 | Improved |
| add EPS to multi-similarity loss | MS · Cars-196 | Recall@1: 81.2 -> 82.9 | Improved |
| add EPS to multi-similarity loss | MS · CUB-200-2011 | Recall@1: 62.3 -> 63.3 | Improved |
| add EPS to triplet loss with semi-hard sampling | Trip. + SH · Cars-196 | Recall@1: 76.1 -> 78.3 | Improved |
| add EPS to triplet loss with semi-hard sampling | Trip. + SH · CUB-200-2011 | Recall@1: 61.5 -> 61.8 | No clear effect |
| add EPS to triplet loss with semi-hard sampling | Trip. + SH · Omniglot-letters | Recall@1: 49.4 -> 68.4 | Improved |
| add EPS to multi-similarity loss | MS · Omniglot-languages | Recall@1: 78.8 -> 86.2 | Improved |
| add EPS to semi-hard sampling | Triplet · Omniglot training set, languages | Recall@1: 99.9 -> 94.5 | Hurt |
| add EPS to semi-hard sampling | Triplet · Omniglot training set, letters | Recall@1: 70.3 -> 77.5 | Improved |
| add EPS to MS with 512-d embedding and Inception backbone | MS · Cars-196 | Recall@1: 84.1 -> 85.5 | Improved |
Closest papers by meaning
- Hard negative examples are hard, but useful (ECCV 2020 · similarity 0.90)
- Learning Towards The Largest Margins (ICLR 2022 · similarity 0.89)
- MIC: Mining Interclass Characteristics for Improved Metric Learning (ICCV 2019 · similarity 0.89)
- Deep Metric Learning With Tuplet Margin Loss (ICCV 2019 · similarity 0.89)
- Supervised Metric Learning to Rank for Retrieval via Contextual Similarity Optimization (ICML 2023 · similarity 0.89)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · similarity 0.88)
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.
