Research map / Label-efficient and robust learning / Recognition and long-tail
Class-Difficulty Based Methods for Long-Tailed Visual Recognition
Saptarshi Sinha, Hiroki Ohashi, Katsuyuki Nakamura
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 |
|---|---|---|---|
| add CDB class-difficulty weighting to focal loss | Focal loss · ImageNet-LT (imbalance 256) | classification error (%): 69.50 -> 66.30 | Improved |
| replace class-frequency sampling with CDB-S | CFS · CIFAR100-LT (imbalance 100) | classification error (%): 61.07 -> 58.03 | Improved |
| use sigmoidal instead of linear dynamic tau (ablation) | CDB-W-CE · CIFAR100-LT (imbalance 100) | classification error (%): 58.40 -> 57.41 | Improved |
| raise fixed tau from 1.5 to 5.0 (ablation) | CDB-W-CE · CIFAR100-LT (imbalance 100) | classification error (%): 57.30 -> 59.38 | Hurt |
| use sigmoidal dynamic tau instead of best fixed tau (ablation) | CDB-W-CE · CIFAR100-LT (imbalance 100) | classification error (%): 57.30 -> 57.41 | No clear effect |
| use class-level CDB-W-CE instead of instance-level focal loss | Focal loss · CIFAR100-LT (imbalance 100) (imbalance 200 analysis) | classification accuracy: 5.11% and 7.26% on medium-shot and few-shot; overall 1.74% | Improved |
| use CDB-S rather than class-aware sampling in cRT stage 2 (CE stage 1) (ablation) | cRT decoupled learning · CIFAR100-LT (imbalance 100) | classification error (%): 57.06 -> 56.49 | Improved |
| use CDB-W-CE weighting rather than class-aware sampling in cRT stage 2 (CE stage 1) (ablation) | cRT decoupled learning · CIFAR100-LT (imbalance 100) | classification error (%): 57.06 -> 57.44 | Hurt |
| use CDB-W-CE instead of CE in stage 1 with CDB-S stage 2 (cRT) (ablation) | cRT decoupled learning · CIFAR100-LT (imbalance 100) | classification error (%): 56.49 -> 56.24 | Improved |
| train detector with CDB-W-CE instead of CE | Faster R-CNN · LVIS v0.5 val (detection) | mAP / APr: 21.19 -> 23.37 mAP; 3.05 -> 9.83 APr | Improved |
| initialize BAGS-cls from CDB-W-CE pretrained model | BAGS-cls · LVIS v0.5 val (detection) | mAP / APr: 25.96 -> 26.01 mAP; 17.65 -> 15.47 APr | Mixed |
| initialize LRW-cls from CDB-W-CE pretrained model | LRW-cls · LVIS v0.5 val (detection) | mAP: 24.66 -> 25.52 | Improved |
Closest papers by meaning
- Deep Long-Tailed Learning: A Survey (TPAMI 2023 · similarity 0.91)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.90)
- Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · similarity 0.90)
- Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition From a Domain Adaptation Perspective (CVPR 2020 · similarity 0.89)
- Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.88)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (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.
