Research map / Label-efficient and robust learning / Recognition and long-tail
Asymmetric Loss for Multi-Label Classification
Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, Lihi Zelnik-Manor
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 cross-entropy with focal loss | TResNet-L · MS-COCO | mAP: 84.0 -> 85.1 | Improved |
| add asymmetric focusing (lower gamma+, gamma+=0) (ablation) | focal loss · MS-COCO | mAP: 85.1 -> 86.3 | Improved |
| add asymmetric probability margin (ablation) | focal loss · MS-COCO | mAP: 85.1 -> 86.3 | Improved |
| add probability margin to asymmetric focusing (ablation) | ASL · MS-COCO | mAP: 86.3 -> 86.6 | Improved |
| use ASL instead of static linear weighting (ablation) | focal loss · MS-COCO | mAP: 85.3 -> 86.6 | Improved |
| add linear positive weighting (ablation) | focal loss · MS-COCO | mAP: 85.1 -> 85.3 | No clear effect |
| adaptive asymmetry with unbiased target gap 0 (ablation) | ASL · MS-COCO | mAP: 85.1 -> 85.8 | Improved |
| raise target probability gap from 0 to 0.2 (ablation) | adaptive ASL · MS-COCO | mAP: 85.8 -> 86.4 | Improved |
| use adaptive instead of fixed gamma- (ablation) | ASL · MS-COCO | mAP: 86.6 -> 86.4 | Hurt |
| use ImageNet-21K instead of 1K pretraining | ASL TResNet-L · MS-COCO | mAP: 86.6 -> 88.4 | Improved |
| increase input resolution from 448 to 640 | ASL TResNet-L · MS-COCO | mAP: 88.4 -> 89.8 | Improved |
| use TResNet-L instead of ResNet101 backbone | ASL · MS-COCO | mAP: 85.0 -> 86.6 | Improved |
Closest papers by meaning
- Distribution-Balanced Loss for Multi-Label Classification in Long-Tailed Datasets (ECCV 2020 · similarity 0.87)
- Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.86)
- Acknowledging the Unknown for Multi-Label Learning with Single Positive Labels (ECCV 2022 · similarity 0.85)
- Joint Asymmetric Loss for Learning with Noisy Labels (ICCV 2025 · similarity 0.85)
- Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.84)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.84)
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.
