Research map / Label-efficient and robust learning / Recognition and long-tail
ImageNet-Hard: The Hardest Images Remaining from a Study of the Power of Zoom and Spatial Biases in Image Classification
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 |
|---|---|---|---|
| Replace MEMO's AugMix augmentations with RandomResizedCrop zoom-in augmentations | ResNet-50 with MEMO · ImageNet | top-1 accuracy: 76.13 -> 77.50 | Helped |
| Replace MEMO's AugMix augmentations with RandomResizedCrop zoom-in augmentations | ResNet-50 with MEMO · ImageNet-A | top-1 accuracy: 0.00 -> 1.31 | Helped |
| Replace MEMO's AugMix augmentations with RandomResizedCrop zoom-in augmentations | ResNet-50 with MEMO · ImageNet-R | top-1 accuracy: 36.17 -> 40.81; 41.28 -> 40.81 | Mixed |
| Replace MEMO's AugMix augmentations with RandomResizedCrop zoom-in augmentations | ResNet-50 with MEMO · ImageNet-Sketch | top-1 accuracy: 24.09 -> 27.53; 27.63 -> 27.53 | Mixed |
| Replace MEMO's AugMix augmentations with RandomResizedCrop zoom-in augmentations | ResNet-50 with MEMO · ObjectNet | top-1 accuracy: 35.92 -> 38.85 | Helped |
| Use zoom-in RandomResizedCrop transforms instead of AugMix within MEMO | Three ResNet-50 variants across five benchmarks · Mean across five benchmarks | top-1 accuracy: 39.26 -> 39.54; +0.28 points | Helped |
| Use zoom-in RandomResizedCrop transforms instead of AugMix within MEMO | Three ResNet-50 variants · ImageNet-A | top-1 accuracy: +1.10 points | Helped |
| Aggregate zoom-in predictions with max rather than use the 1-crop prediction | ViT-B/32 · ImageNet-A | top-1 accuracy: 24.69%; +15.05 | Helped |
| Aggregate zoom-in predictions with max rather than use the 1-crop prediction | ResNet-50 · ImageNet-A | top-1 accuracy: 16.62% -> 29.65%; +13.03 | Helped |
| Aggregate zoom-in predictions with mean rather than use the 1-crop prediction | Image classifiers · ImageNet, ReaL, ImageNet-A, ObjectNet | top-1 accuracy: +3.56 on IN; +4.08 on ReaL; +4.65 on IN-A; +3.03 on ON | Helped |
| Use zoom-in transforms in simple mean/max aggregation | Image classifiers · ImageNet-R and ImageNet-Sketch | top-1 accuracy | Hurt |
| Use zoom-out transforms for abstract images | Image classifiers · ImageNet-R and ImageNet-Sketch | top-1 accuracy: 1.2% - 3% | Helped |
Closest papers by meaning
- Hard ImageNet: Segmentations for Objects with Strong Spurious Cues (NeurIPS 2022 · similarity 0.90)
- When does dough become a bagel? Analyzing the remaining mistakes on ImageNet (NeurIPS 2022 · similarity 0.89)
- Doomed to Re-Annotate, Forever: The ImageNet Story (NeurIPS 2026 · similarity 0.89)
- SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification (NeurIPS 2024 · similarity 0.88)
- ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models (NeurIPS 2019 · similarity 0.88)
- From ImageNet to Image Classification: Contextualizing Progress on Benchmarks (ICML 2020 · 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.