Research map / Label-efficient and robust learning / Recognition and long-tail
Long-Tail Recognition via Compositional Knowledge Transfer
Sarah Parisot, Pedro M. Esperança, Steven McDonagh, Tamas J. Madarasz, Yongxin Yang, Zhenguo Li
Check 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 |
|---|---|---|---|
| Compose prototype-anchored attention over learned common-class classifiers and ensemble classifiers tuned for different class groups | Long-tail classifier on ImageNet-LT · ImageNet-LT | Few-shot accuracy: 20.6 -> 36.9 | Improved |
| Compose prototype-anchored attention over classifiers and ensemble class-group-specific classifiers | Long-tail classifier on ImageNet-LT · ImageNet-LT | Total accuracy: 50.9 -> 54.2 | Improved |
| Use a single k=20 hybrid classifier rather than the ensemble (ablation) | Compositional knowledge-transfer classifier · ImageNet-LT | Few-shot accuracy: 48.2 vs 36.9 | Mixed |
| Add the cosine classifier rather than using prototypes alone (ablation) | Prototype classifier · ImageNet-LT | Few-shot accuracy: 27.4 -> 25.2 | Hurt |
| Exclude few-shot classes from training and estimate their classifiers from prototypes and common-class weights (ablation) | Continual classifier · ImageNet-LT, few-shot classes excluded from training | Few-shot accuracy: 26.2 -> 32.2 | Improved |
| Reverse prototype-to-classifier attention to classifier-to-prototype attention (ablation) | Compositional knowledge-transfer classifier · ImageNet-LT | Few-shot accuracy: 36.9 -> 28.9 | Hurt |
| Replace prototype-to-classifier attention with classifier self-attention (ablation) | Compositional knowledge-transfer classifier · ImageNet-LT | Few-shot accuracy: 36.9 -> 31.3 | Hurt |
| Replace prototype-to-classifier attention with prototype self-attention (ablation) | Compositional knowledge-transfer classifier · ImageNet-LT | Few-shot accuracy: 36.9 -> 31.4 | Hurt |
| Train with a cosine rather than a dot-product classifier (ablation) | ImageNet-LT backbone · ImageNet-LT | Few-shot classifier accuracy: 7.5 -> 15.4 | Improved |
| Use square-root rather than uniform sampling (ablation) | ImageNet-LT backbone training · ImageNet-LT | Few-shot classifier accuracy: 15.4 -> 20.6 | Improved |
| Use class-aware rather than square-root sampling (ablation) | ImageNet-LT backbone training · ImageNet-LT | Few-shot classifier accuracy: 20.6 -> 14.2 | Hurt |
| Use balanced softmax rather than cross-entropy for backbone training (ablation) | ImageNet-LT compositional classifier ensemble · ImageNet-LT | Few-shot accuracy: 36.9 -> 38.8 | Improved |
Closest papers by meaning
- Imagine by Reasoning: A Reasoning-Based Implicit Semantic Data Augmentation for Long-Tailed Classification (AAAI 2022 · similarity 0.88)
- Echoes of the Past: Boosting Long-tail Recognition via Reflective Learning (ECCV 2024 · similarity 0.88)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.88)
- Learning from Neighbors: Category Extrapolation for Long-Tail Learning (CVPR 2025 · similarity 0.86)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.86)
- LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content (CVPR 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.
