AI Paper Atlas

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

CVPR 2022 · Label-efficient and robust learning › Recognition and long-tail · approach: Decoupled and calibrated long-tail training · 33 citations (OpenAlex)

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 changedWhereMeasuredEffect
Compose prototype-anchored attention over learned common-class classifiers and ensemble classifiers tuned for different class groupsLong-tail classifier on ImageNet-LT · ImageNet-LTFew-shot accuracy: 20.6 -> 36.9Improved
Compose prototype-anchored attention over classifiers and ensemble class-group-specific classifiersLong-tail classifier on ImageNet-LT · ImageNet-LTTotal accuracy: 50.9 -> 54.2Improved
Use a single k=20 hybrid classifier rather than the ensemble (ablation)Compositional knowledge-transfer classifier · ImageNet-LTFew-shot accuracy: 48.2 vs 36.9Mixed
Add the cosine classifier rather than using prototypes alone (ablation)Prototype classifier · ImageNet-LTFew-shot accuracy: 27.4 -> 25.2Hurt
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 trainingFew-shot accuracy: 26.2 -> 32.2Improved
Reverse prototype-to-classifier attention to classifier-to-prototype attention (ablation)Compositional knowledge-transfer classifier · ImageNet-LTFew-shot accuracy: 36.9 -> 28.9Hurt
Replace prototype-to-classifier attention with classifier self-attention (ablation)Compositional knowledge-transfer classifier · ImageNet-LTFew-shot accuracy: 36.9 -> 31.3Hurt
Replace prototype-to-classifier attention with prototype self-attention (ablation)Compositional knowledge-transfer classifier · ImageNet-LTFew-shot accuracy: 36.9 -> 31.4Hurt
Train with a cosine rather than a dot-product classifier (ablation)ImageNet-LT backbone · ImageNet-LTFew-shot classifier accuracy: 7.5 -> 15.4Improved
Use square-root rather than uniform sampling (ablation)ImageNet-LT backbone training · ImageNet-LTFew-shot classifier accuracy: 15.4 -> 20.6Improved
Use class-aware rather than square-root sampling (ablation)ImageNet-LT backbone training · ImageNet-LTFew-shot classifier accuracy: 20.6 -> 14.2Hurt
Use balanced softmax rather than cross-entropy for backbone training (ablation)ImageNet-LT compositional classifier ensemble · ImageNet-LTFew-shot accuracy: 36.9 -> 38.8Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy