Research map / Label-efficient and robust learning / Test-time adaptation
Fixing the train-test resolution discrepancy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| increase test resolution from 224 to 288 without fine-tuning | ResNet-50 trained at 224 · ImageNet val | top-1 accuracy: 77.0 -> 78.4 | Helped |
| increase test resolution further to 448 without fine-tuning | ResNet-50 trained at 224 · ImageNet val | top-1 accuracy: 78.4 -> 76.6 | Hurt |
| fine-tune classifier and last BN at higher test resolution | ResNet-50 trained at 224 · ImageNet val | top-1 accuracy: 78.4 -> 79.1 | Helped |
| fine-tune classifier only (FC) at test resolution (ablation) | ResNet-50 trained at 224 · ImageNet val | top-1 accuracy: 78.9 | Helped |
| add batch-norm fine-tuning and better augmentation (ablation) | ResNet-50 trained at 224 · ImageNet val | top-1 accuracy: 78.9 -> 79.0 | No clear effect |
| train at 128 and fine-tune for test 224 instead of train at 224 | ResNet-50 · ImageNet val | top-1 accuracy: 77.0 -> 77.1 | Helped |
| fine-tune BN and classifier at higher test resolution (ablation) | PNASNet-5-Large · ImageNet val | top-1 accuracy: 82.8 -> 83.5 | Helped |
| also fine-tune last three cells (ablation) | PNASNet-5-Large · ImageNet val | top-1 accuracy: 83.4 -> 83.7 | Helped |
| fine-tune and test at 320 instead of 224 | ResNeXt-101 32x48d · ImageNet val | top-1 accuracy: 85.4 -> 86.4 | Helped |
| apply resolution adaptation for transfer | SENet-154 / InceptionResNet-V2 · Birdsnap | top-1 accuracy: 83.4 -> 84.3 | Helped |
| apply resolution adaptation for transfer | SENet-154 · Stanford Cars | top-1 accuracy: 94.0 -> 94.4 | Helped |
| parametric Frechet equalization after pooling instead of fine-tuning (ablation) | ResNet-50 · ImageNet val | top-1 accuracy | No clear effect |
Closest papers by meaning
- Learning Loss for Test-Time Augmentation (NeurIPS 2020 · similarity 0.85)
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.85)
- Understanding the detrimental class-level effects of data augmentation (NeurIPS 2023 · similarity 0.85)
- Learning Augmentation Network via Influence Functions (CVPR 2020 · similarity 0.84)
- Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation (NeurIPS 2026 · similarity 0.82)
- Not Just Pretty Pictures: Toward Interventional Data Augmentation Using Text-to-Image Generators (ICML 2024 · similarity 0.82)
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.