Research map / Label-efficient and robust learning / Recognition and long-tail
Delving Deep into Simplicity Bias for Long-Tailed Image Recognition
Xiu-Shen Wei, Xuhao Sun, Yang Shen, Peng Wang
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 holistic-level SSL to the CE baseline (ablation) | CE baseline · ImageNet-LT | All accuracy: 52.5 -> 55.8 | Improved |
| add partial-level SSL to the CE baseline (ablation) | CE baseline · ImageNet-LT | All accuracy: 52.5 -> 55.2 | Improved |
| add partial-level SSL on top of holistic-level SSL (ablation) | CE + H-SSL · ImageNet-LT | All accuracy: 55.8 -> 57.0 | Improved |
| add augmented-level SSL on top of holistic and partial levels (ablation) | CE + H-SSL + P-SSL · ImageNet-LT | All accuracy: 57.0 -> 59.1 | Improved |
| add augmented-level SSL on top of holistic and partial levels (ablation) | CE + H-SSL + P-SSL · ImageNet-LT (Few) | Few accuracy: 35.7 -> 38.3 | Improved |
| use all queue examples of predicted class c' instead of a single example (ablation) | 3LSSL augmented queue · ImageNet-LT | All accuracy: 57.8 -> 59.1 | Improved |
| use predicted class c' rather than ground-truth class c for the augmented queue (ablation) | 3LSSL augmented queue · ImageNet-LT | All accuracy: 58.9 -> 59.1 | Improved |
| use predicted class c' rather than ground-truth class c for the augmented queue (ablation) | 3LSSL augmented queue · ImageNet-LT (Few) | Few accuracy: 37.8 -> 38.3 | Improved |
| increase partial-level masking stages from 1 to 3 (ablation) | 3LSSL partial-level SSL · ImageNet-LT | All accuracy: 59.1 -> 58.7 | Hurt |
| increase partial-level masking stages from 1 to 2 (ablation) | 3LSSL partial-level SSL · ImageNet-LT | All accuracy: 59.1 -> 59.0 | No clear effect |
| add MoCo-v2 SSL to CE pre-training | CE (frozen backbone, classifier re-trained) · Long-tailed CIFAR-100 (IF 100) | Tail accuracy: 45.36 -> 48.20 | Improved |
| use SSL alone instead of CE for pre-training | frozen backbone, classifier re-trained · Long-tailed CIFAR-100 (IF 100) | All accuracy: 55.50 -> 28.89 | Hurt |
Closest papers by meaning
- Constructing Balance from Imbalance for Long-Tailed Image Recognition (ECCV 2022 · similarity 0.90)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.88)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.88)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.87)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · 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.
