Research map / Label-efficient and robust learning / Recognition and long-tail
SuperDisco: Super-Class Discovery Improves Visual Recognition for the Long-Tail
Yingjun Du, Jiayi Shen, Xiantong Zhen, Cees G. M. Snoek
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 the discovered super-class graph and feature-refinement message passing | fine-tuning baseline · CIFAR-100-LT | accuracy (%): 39.3 -> 45.2 | Improved |
| add prototype-graph meta-learning to SuperDisco (ablation) | SuperDisco · CIFAR-100-LT | accuracy (%): 60.3 -> 68.5; 57.3 -> 63.1; 47.5 -> 58.3; 44.9 -> 53.8; 39.3 -> 47.5 | Improved |
| use a multi-level discovered super-class graph rather than one level (20) (ablation) | Meta-SuperDisco · CIFAR-100-LT, imbalance ratio 10 | accuracy (%): 61.2 -> 69.1 | Improved |
| increase hierarchy from (4, 8, 16) to (4, 8, 16, 32) (ablation) | Meta-SuperDisco · CIFAR-100-LT, imbalance ratio 10 | accuracy (%): 69.1 -> 68.5 | No clear effect |
| use five super-class levels (4, 8, 16, 32, 64) instead of four (4, 8, 16, 32) (ablation) | Meta-SuperDisco · CIFAR-100-LT, imbalance ratio 10 | accuracy (%): 68.5 -> 66.9 | Hurt |
| use discovered super-classes rather than ground-truth oracle super-classes (ablation) | Meta-SuperDisco · CIFAR-100-LT, imbalance ratio 10 | accuracy (%): 66.9 -> 69.1 | Improved |
| use a deeper, wider hierarchy rather than one level with 13 super-classes (ablation) | Meta-SuperDisco · iNaturalist, All | accuracy (%): 70.8 -> 72.3 | Improved |
| replace the super-class mechanism with a multi-layer perceptron (ablation) | SuperDisco · ImageNet-LT, All | accuracy (%): 57.1 -> 55.0 | Hurt |
| add the super-class mechanism to the baseline representation (ablation) | ImageNet-LT · ImageNet-LT, All | accuracy (%): 52.7 -> 55.0 | Improved |
| add Meta-SuperDisco to the long-tailed recognition system | long-tailed classifiers · ImageNet-LT; Places-LT; iNaturalist | All accuracy (%): 56.2 -> 57.1; 39.2 -> 40.3; 72.8 -> 73.6 | Mixed |
| add Meta-SuperDisco to the long-tailed recognition system | long-tailed classifiers · ImageNet-LT, Few | accuracy (%): 41.5 -> 37.1 | Hurt |
| use Meta-SuperDisco for long-tailed recognition | long-tailed classifiers · CIFAR-100-LT | accuracy (%): 68.8 -> 69.3; 57.6 -> 58.3; 52.0 -> 53.8 | Improved |
Closest papers by meaning
- Large-Scale Long-Tailed Recognition in an Open World (CVPR 2019 · similarity 0.87)
- Deep Long-Tailed Learning: A Survey (TPAMI 2023 · similarity 0.87)
- Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · similarity 0.87)
- Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition From a Domain Adaptation Perspective (CVPR 2020 · similarity 0.86)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.86)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.84)
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.
