Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Dataset Bias in Few-Shot Image Recognition
Shuqiang Jiang, Yaohui Zhu, Chenlong Liu, Xinhang Song, Xiangyang Li, Weiqing Min
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 |
|---|---|---|---|
| use ResNet-12 backbone instead of CNN-4 | ProtoNet (PN) · five Mini datasets, 5-way 1-shot | accuracy gain: 8.83% | Improved |
| use ResNet-12 backbone instead of CNN-4 | Relation Network (RN) · five Mini datasets, 5-way 1-shot | accuracy gain: 3.53% | Improved |
| use ResNet-12 instead of CNN-4 | all methods · MiniCharacter | average accuracy gain: 4.21% | Improved |
| use ResNet-12 instead of CNN-4 | all methods · MiniPlaces | average accuracy gain: 8.53% | Improved |
| lower image complexity of the dataset (MiniCharacter vs MiniImagenet) (ablation) | PN · 5-way 1-shot, CNN-4 | accuracy (%): 49.42 -> 86.17 | Improved |
| MiniPlaces instead of MiniFood (similar complexity, larger intra-inter consistency gap) (ablation) | all methods · 5-way 1-shot, CNN-4 and ResNet-12 | average accuracy gain: 6.8% (CNN-4), 7.8% (ResNet-12) | Improved |
| use DEMD instead of PN on fine-grained MiniFood (ablation) | DEMD · MiniFood 5-way 1-shot, CNN-4 | accuracy (%): 41.41 -> 50.70 | Improved |
| use MetaOpt instead of PN on MiniPlaces (ablation) | MetaOpt · MiniPlaces 5-way 1-shot, CNN-4 | accuracy (%): 49.00 -> 55.74 | Improved |
| initialize classifier with prototypes (Proto-MAML instead of MAML) (ablation) | Proto-MAML · MiniCharacter, ResNet-12 | accuracy (%): 85.65 -> 94.24 | Improved |
| more instances per base category (10 to 1,200) with relevant novel categories (ablation) | base categories (LTED) · SG-(L;L) 5-way 1-shot | accuracy: improvement fast from 10 to 200 samples, slower after | Improved |
| more instances per base category with irrelevant novel categories (ablation) | base categories (LTED) · SG-(L;AI) 5-way 1-shot | accuracy: no significant gain after 100 samples per category | Mixed |
| add irrelevant additional base categories (AROT) to LTED base for LTED novel (ablation) | additional base categories · CG-(L,AO;L) 5-way 1-shot | accuracy: improves at first, then drops | Mixed |
Closest papers by meaning
- Impact of base dataset design on few-shot image classification (ECCV 2020 · similarity 0.90)
- Benchmarking Spurious Bias in Few-Shot Image Classifiers (ECCV 2024 · similarity 0.90)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.90)
- LibFewShot: A Comprehensive Library for Few-Shot Learning (TPAMI 2023 · similarity 0.89)
- Few-Shot Image Recognition by Predicting Parameters From Activations (CVPR 2018 · similarity 0.89)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.88)
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.
