Research map / Label-efficient and robust learning / Recognition and long-tail
Stochastic Class-Based Hard Example Mining for Deep Metric Learning
Check 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 |
|---|---|---|---|
| Use stochastic class-based hard negative mining instead of random batch sampling | Triplet-loss metric learning · CARS-196, CUB-200-2011, In-shop retrieval, Stanford Online Products | Recall@K | Helped |
| Use stochastic class-based hard negative mining instead of class-level mining | Triplet-loss metric learning · CARS-196, CUB-200-2011, In-shop retrieval, Stanford Online Products | number of triplets with non-zero loss | Helped |
| Use second-order pooling with cross-channel correlation instead of Inception v1 features | Image retrieval · CARS-196, 224 × 224 | R@1 (%): 83.6 -> 86.9 | Helped |
| Use second-order pooling with cross-channel correlation instead of Inception v1 features | Image retrieval · CARS-196, 336 × 336 | R@1 (%): 89.7 -> 91.3 | Helped |
| Use second-order pooling with cross-channel correlation instead of Inception v1 features | Image retrieval · CUB-200-2011, 224 × 224 | R@1 (%): 55.1 -> 58.1 | Helped |
| Use second-order pooling with cross-channel correlation instead of Inception v1 features | Image retrieval · CUB-200-2011, 336 × 336 | R@1 (%): 60.9 -> 65.2 | Helped |
| Increase input resolution from 224 × 224 to 336 × 336 (ablation) | Inception v1 feature extractor · CARS-196 | R@1 (%): 83.6 -> 89.7 | Helped |
| Increase input resolution from 224 × 224 to 336 × 336 (ablation) | Channel-correlation feature extractor · CARS-196 | R@1 (%): 86.9 -> 91.3 | Helped |
| Increase input resolution from 224 × 224 to 336 × 336 (ablation) | Inception v1 feature extractor · CUB-200-2011 | R@1 (%): 55.1 -> 60.9 | Helped |
| Increase input resolution from 224 × 224 to 336 × 336 (ablation) | Channel-correlation feature extractor · CUB-200-2011 | R@1 (%): 58.1 -> 65.2 | Helped |
| Allow class-signature loss to back-propagate to the feature extractor (ablation) | Stochastic hard example mining, full model · CARS-196, CUB-200-2011, In-shop retrieval, Stanford Online Products | Recall@K | Helped |
| Replace learned class signatures with precomputed class-wise average features (ablation) | Stochastic hard example mining · CARS-196, CUB-200-2011, In-shop retrieval, Stanford Online Products | Recall@K | Hurt |
Closest papers by meaning
- Deep Metric Learning by Online Soft Mining and Class-Aware Attention (AAAI 2019 · similarity 0.90)
- Symmetrical Synthesis for Deep Metric Learning (AAAI 2020 · similarity 0.87)
- Noise-Resistant Deep Metric Learning With Ranking-Based Instance Selection (CVPR 2021 · similarity 0.87)
- Signal-To-Noise Ratio: A Robust Distance Metric for Deep Metric Learning (CVPR 2019 · similarity 0.87)
- Deep Metric Learning via Lifted Structured Feature Embedding (CVPR 2016 · similarity 0.87)
- Mathematical Justification of Hard Negative Mining via Isometric Approximation Theorem (ICLR 2024 · 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.