Research map / Label-efficient and robust learning / Recognition and long-tail
Difficulty-aware Balancing Margin Loss for Long-tailed Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes difficulty-aware balancing margin (DBM) loss, combining a class-wise frequency margin with an instance-wise angular-difficulty margin applied to hard positives, and plugs it into existing long-tailed recognition methods to improve accuracy on CIFAR-LT, ImageNet-LT and iNaturalist2018.
Problem
Deep networks trained on long-tailed data perform poorly on minority classes; prior rebalancing methods address class-level imbalance but ignore instance-level difficulty variation within each class.
Approach
A margin loss on a cosine classifier with a class-wise margin scaled inversely with class frequency plus an instance-wise margin proportional to the angular difficulty of misclassified (hard positive) samples, which is compatible with CE, CB/DRW, BS, BCL, GML and NCL.
Key results
- DBM-BS reaches 51.30% top-1 on CIFAR-100-LT (IF 100) vs 49.35% for BS, and DBM-NCL reaches 57.48% vs 56.68% for NCL.
- DBM-GML achieves 57.4% on ImageNet-LT with ResNet-50 (vs 57.2% reported for GML) and 72.0% on iNaturalist2018 with ResNet-50 (vs 71.2% for reproduced GML).
- Ablation on CIFAR-100-LT shows the instance-wise margin adds about 0.7 points to both CE and BS, and applying it only to hard positives beats applying it to all positives.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DBM loss to CE | CE · CIFAR-10-LT (IF 100) | top-1 accuracy: 78.48 -> 80.84 | Helped |
| Add DBM loss to CE | CE · CIFAR-100-LT (IF 100) | top-1 accuracy: 44.60 -> 46.53 | Helped |
| Add DBM loss to BCL | BCL · ImageNet-LT, ResNet-50 and ResNeXt-50 | top-1 accuracy: 56.0 -> 56.3; 56.7 -> 57.4 | Helped |
| Add DBM loss to GML | GML · ImageNet-LT, ResNeXt-50 | top-1 accuracy: 58.3 -> 58.6 | Helped |
| Add DBM loss to BCL or GML | BCL and GML · iNaturalist2018, overall | top-1 accuracy: +0.9%p; +0.8%p | Helped |
| Replace the linear classifier with a cosine classifier only (ablation) | CE and BS · CIFAR-100-LT (IF 100) | top-1 accuracy: 44.60 -> 44.29; 49.35 -> 49.84 | No clear effect |
| Add a class-wise margin (ablation) | Cosine-classifier CE and BS · CIFAR-100-LT (IF 100) | top-1 accuracy: 44.29 -> 45.85; 49.84 -> 50.61 | Helped |
| Add instance-wise margin to all positive samples (ablation) | Cosine + class-wise margin CE and BS · CIFAR-100-LT (IF 100) | top-1 accuracy: 45.85 -> 46.38; 50.61 -> 50.93 | Helped |
| Apply instance-wise margin only to hard positive samples instead of all positive samples (ablation) | Cosine + class-wise margin CE and BS · CIFAR-100-LT (IF 100) | top-1 accuracy: 46.38 -> 46.53; 50.93 -> 51.30 | Helped |
| Add DBM loss to BS | BS · CIFAR-100-LT (IF 100), inter-class separability overall | Fisher criterion: 5.89 -> 6.15 | Helped |
Surprising findings
- Using a cosine classifier alone does not significantly improve performance over the linear classifier (44.29 vs 44.60 for CE).
Limitations
- Instance-wise margin only applied to hard positives identified as misclassified during training, a heuristic choice.
- Gains are modest on several benchmarks (e.g., ~0.3 points on ImageNet-LT with BCL ResNet-50, marginal overall gain with CLIP on ImageNet-LT).
- Hyperparameter K is tuned per dataset and baseline within 0.1-0.3.
Experimental setup
Datasets: CIFAR-10-LT CIFAR-100-LT ImageNet-LT iNaturalist2018
Benchmarks: CIFAR-10-LT CIFAR-100-LT ImageNet-LT iNaturalist2018
Metrics: top-1 accuracy Many/Medium/Few group accuracy Fisher criterion inter-class separability mean angular distance
Baselines: CE LDAM-DRW CE-DRW BS BCL GML NCL RIDE DisAlign GLMC Yu et al. (2022)
Builds on: LDAM ArcFace balanced softmax deferred re-weighting BCL NCL GML
Closest papers by meaning
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.91)
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.89)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.88)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.88)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.87)
- Inverse Weight-Balancing for Deep Long-Tailed Learning (AAAI 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.