Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Impact of base dataset design on few-shot image classification
Othman Sbai, Camille Couprie, Mathieu Aubry
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 |
|---|---|---|---|
| train on full MiniIN6K (6000 classes) instead of the 64 MiniIN base classes | Cosine Classifier, WRN28-10 · MiniIN test 5-shot | accuracy: 78.95 -> 96.91 | Improved |
| train on full MiniIN6K instead of MiniIN base classes | Cosine Classifier, WRN28-10 · MiniIN test 1-shot | accuracy: 61.62 -> 85.40 | Improved |
| use random 64 classes from MiniIN6K instead of the MiniIN benchmark base classes | Cosine Classifier, WRN28-10 · MiniIN test 5-shot | accuracy: 78.95 -> 75.48 | Hurt |
| train on full MiniIN6K* instead of CUB base classes | Cosine Classifier, WRN28-10 · CUB test 5-shot | accuracy: 90.32 -> 90.89 | No clear effect |
| train ProtoNet on full MiniIN6K instead of MiniIN base | ProtoNet, WRN28-10 · MiniIN test 5-shot | accuracy: 73.64 -> 85.14 | Improved |
| train on IN6k instead of the 389 ImageNet base classes | Cosine Classifier, ResNet-34 · ImageNet benchmark 5-shot | top-5 accuracy: 78.15 -> 88.50 | Improved |
| use about 384 classes with 100 images per class instead of 64 classes with 600 images for a fixed budget (ablation) | Cosine Classifier · MiniIN test 5-shot | accuracy: around 5% | Improved |
| split base classes into 8 sub-classes using oracle features (ablation) | Cosine Classifier, ResNet-34 · ImageNet benchmark 5-shot | top-5 accuracy: 78.15 -> 80.30 | Improved |
| split base classes into 4 sub-classes using oracle features (ablation) | Cosine Classifier, ResNet-34 · ImageNet benchmark 1-shot | top-5 accuracy: 57.56 -> 58.49 | Improved |
| select base classes closest rather than farthest to the test classes (ablation) | Cosine Classifier, WRN28-10 · MiniIN test 5-shot | accuracy | Improved |
| regroup or split classes toward the optimal class count (grouping if too many classes, splitting if too few) (ablation) | Cosine Classifier · MiniIN and CUB test | accuracy | Mixed |
| use balanced bisection/pairing relabeling instead of K-means (ablation) | Cosine Classifier · MiniIN test | accuracy | Improved |
Closest papers by meaning
- Learning to Select Base Classes for Few-Shot Classification (CVPR 2020 · similarity 0.91)
- A Closer Look at Few-shot Classification (ICLR 2019 · similarity 0.90)
- Dataset Bias in Few-Shot Image Recognition (TPAMI 2022 · similarity 0.90)
- A Baseline for Few-Shot Image Classification (ICLR 2020 · similarity 0.90)
- Few-Class Arena: A Benchmark for Efficient Selection of Vision Models and Dataset Difficulty Measurement (ICLR 2025 · similarity 0.88)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 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.
