Research map / Label-efficient and robust learning / Recognition and long-tail
Wrapped Cauchy Distributed Angular Softmax for Long-Tailed Visual 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 |
|---|---|---|---|
| replace angular softmax with WCDAS | ResNet-32 · CIFAR-100-LT, imbalance factors 200, 100, and 10 | top-1 accuracy: 44.2 -> 49.3; 49.7 -> 52.5; 64.1 -> 65.8 | Helped |
| replace angular softmax with WCDAS | ResNet-32 · CIFAR-10-LT, imbalance factors 200, 100, and 10 | top-1 accuracy: 80.9 -> 81.7; 83.8 -> 86.4; 91.4 -> 92.4 | Helped |
| replace t-vMF Similarity with WCDAS | ResNet-10 · ImageNet-LT, all classes | top-1 accuracy: 43.5 -> 44.5 | Helped |
| replace t-vMF Similarity with WCDAS | ResNet-50 · iNaturalist 2018, all classes | top-1 accuracy: 71.0 -> 71.8 | Helped |
| use wrapped Cauchy instead of wrapped Normal (ablation) | WCDAS, ImageNet-LT · ImageNet-LT, all classes | top-1 accuracy: 42.1 -> 43.8 (shared parameter); 42.3 -> 44.5 (class-wise parameters) | Helped |
| use class-wise rather than shared concentration parameters (ablation) | WCDAS, ImageNet-LT · ImageNet-LT, all classes | top-1 accuracy: 43.8 -> 44.5 | Helped |
| use alternative initializations for wρ instead of init -1.0 (ablation) | WCDAS, ImageNet-LT · ImageNet-LT, all classes | top-1 accuracy: 44.5; 43.9; 44.3; 44.2; 44.3; 44.2; 44.0 | Mixed |
| extend training from 90 to 150 epochs (ablation) | WCDAS, ImageNet-LT with learning rate 0.2 · ImageNet-LT, all classes | top-1 accuracy: 43.7 -> 44.5 | Helped |
| extend training from 90 to 150 epochs (ablation) | WCDAS, ImageNet-LT with learning rate 0.2 · ImageNet-LT, few-shot classes | top-1 accuracy: 22.8 -> 21.7 | Hurt |
| use meta-sampling instead of class-balanced sampling (ablation) | WCDAS, ImageNet-LT classifier learning · ImageNet-LT, all classes | top-1 accuracy: 44.5 -> 44.0; 44.1; 43.7 | Hurt |
| use meta-sampling instead of class-balanced sampling (ablation) | WCDAS, ImageNet-LT classifier learning · ImageNet-LT, few-shot classes | top-1 accuracy: 24.1 -> 23.0; 25.3; 27.7 | Mixed |
Closest papers by meaning
- Balanced Meta-Softmax for Long-Tailed Visual Recognition (NeurIPS 2020 · similarity 0.91)
- Global and Local Mixture Consistency Cumulative Learning for Long-Tailed Visual Recognitions (CVPR 2023 · similarity 0.87)
- Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition From a Domain Adaptation Perspective (CVPR 2020 · similarity 0.86)
- Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.86)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.85)
- Improving Visual Prompt Tuning by Gaussian Neighborhood Minimization for Long-Tailed Visual Recognition (NeurIPS 2024 · similarity 0.85)
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.