Research map / Label-efficient and robust learning / Recognition and long-tail
Decoupling Representation and Classifier for Long-Tailed Recognition
Read the paperCheck 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 |
|---|---|---|---|
| Decouple classifier learning from representation learning and rebalance the classifier | ResNeXt-50 on ImageNet-LT · ImageNet-LT | top-1 accuracy (All): 44.4 -> 49.6 | Helped |
| Use instance-balanced sampling to learn the representation, then balance the classifier | Decoupled long-tailed recognition · ImageNet-LT | top-1 accuracy (All): 44.4 -> 49.6 | Helped |
| Use progressively-balanced instead of instance-balanced sampling during joint training | Joint ResNeXt-50 · ImageNet-LT | top-1 accuracy (All): 44.4 -> 47.2 | Helped |
| Add classifier weight normalization | Jointly trained classifier · ImageNet-LT | top-1 accuracy (All): 44.4 -> 49.4 | Helped |
| Freeze the representation and retrain only the classifier, rather than fine-tuning backbone and classifier together (ablation) | ResNeXt-50 · ImageNet-LT | top-1 accuracy (All): 46.3 -> 49.5 | Helped |
| Replace the linear classifier with a deeper MLP (ablation) | ResNeXt-50 representation · ImageNet-LT | top-1 accuracy (All): 49.4 -> 48.0; 49.4 -> 47.7 | Hurt |
| Replace the linear classifier with a deeper MLP (ablation) | ResNet-50 representation · iNaturalist 2018 | top-1 accuracy (All): 68.2 -> 61.2; 68.2 -> 62.8 | Hurt |
| Select τ on the training set instead of the validation set (ablation) | τ-normalized classifier · ImageNet-LT | top-1 accuracy (All): 49.4 -> 49.4 | No clear effect |
| Learn τ on the training set rather than select it by grid search (ablation) | τ-normalized classifier · ImageNet-LT | top-1 accuracy (All): 49.4 -> 49.4 | No clear effect |
| Learn τ on the training set instead of use validation-selected τ (ablation) | τ-normalized classifier · Places-LT | top-1 accuracy (All): 37.9 -> 36.8 | Hurt |
| Add the OLTR memory mechanism during cRT classifier retraining (ablation) | cRT · ImageNet-LT | top-1 accuracy | No clear effect |
| Replace linear classification with cosine similarity classification (ablation) | ResNeXt-50 representation · ImageNet-LT | top-1 accuracy (All): 49.4 -> 49.7 | No clear effect |
Closest papers by meaning
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.91)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.91)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.91)
- Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.91)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.90)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.90)
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.