Research map / Label-efficient and robust learning / Recognition and long-tail
Distilling Virtual Examples for Long-Tailed Recognition
Yin-Yin He, Jianxin Wu, Xiu-Shen Wei
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 BSCE loss term to the distillation loss (ablation) | DiVE · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 44.50 -> 45.35 | Improved |
| add power normalization to virtual examples (tau=1, label-mixed teacher) (ablation) | DiVE variant · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 41.59 -> 44.55 | Improved |
| raise temperature from 1 to 3 (on power-normalized label-mixed teacher) (ablation) | DiVE variant · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 44.55 -> 43.22 | Hurt |
| raise temperature to 3 and use power normalization on tau-softened teacher without ground truth (ablation) | DiVE without BSCE · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 43.25 -> 44.50 | Improved |
| replace label-mixed teacher with plain teacher logits (tau=3, power) (ablation) | DiVE without BSCE · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 43.22 -> 44.50 | Mixed |
| distill virtual examples from a model trained on full CIFAR-100 (tau=3) instead of CE teacher | student · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 39.20 -> 61.58 | Improved |
| use a more balanced teacher (BSCE instead of CE) for distillation | student · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 39.20 -> 43.25 | Improved |
| apply DiVE on top of BSCE | BSCE · ImageNet-LT | top-1 accuracy: 50.48 -> 53.10 | Improved |
| apply DiVE on top of BSCE | BSCE · iNaturalist2018 (90 epochs) | top-1 accuracy: 65.35 -> 69.13 | Improved |
| train DiVE for 200 instead of 90 epochs | DiVE · iNaturalist2018 | top-1 accuracy: 69.13 -> 71.71 | Improved |
| distill RIDE with DiVE (RIDE-DiVE) | RIDE · iNaturalist2018 | top-1 accuracy: 72.60 -> 73.44 | Improved |
| distill RIDE with DiVE (RIDE-DiVE) | RIDE · CIFAR-100-LT (imbalance 100) | top-1 accuracy: 49.10 -> 51.66 | Improved |
Closest papers by meaning
- Self Supervision to Distillation for Long-Tailed Visual Recognition (ICCV 2021 · similarity 0.91)
- Bag of Tricks for Long-Tailed Visual Recognition with Deep Convolutional Neural Networks (AAAI 2021 · similarity 0.90)
- A Survey on Long-Tailed Visual Recognition (IJCV 2022 · similarity 0.90)
- DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.89)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 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.
