Research map / Label-efficient and robust learning / Recognition and long-tail
Lost in Translation: Modern Neural Networks Still Struggle With Small Realistic Image Transformations
Ofir Shifman, Yair Weiss
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 |
|---|---|---|---|
| apply RICS crop selection (random-filter scoring) to DINOv2 at inference | DINOv2 · ImageNet subset, 1-pixel realistic shift | Adversarial robustness (1-NN): 62.42 -> 94.73 | Improved |
| apply RICS to DINOv2 on accuracy | DINOv2 · ImageNet subset | accuracy (%): 84.11 -> 82.66 | Hurt |
| use Mexican-hat scoring instead of random filter scoring (ablation) | RICS · ImageNet subset, DINOv2 | Adversarial robustness (1-NN): 94.73 -> 95.14 | No clear effect |
| increase crop size k (ablation) | RICS · ImageNet | accuracy vs robustness | Mixed |
| scale up pretraining data (open-CLIP on LAION-2B, DINOv2) without translation-specific design | open-CLIP, DINOv2 · ImageNet subset, 1-pixel shift | Adversarial robustness (1-NN): 62.42 (DINOv2) vs 46.33 (ResNet50) | No clear effect |
| use cyclic-shift-invariant architecture (AFC) on realistic non-cyclic shifts | AFC · ImageNet subset, 1-pixel realistic shift | Adversarial robustness (1-NN): 100% cyclic vs 89.77 realistic | Mixed |
| apply RICS with cyclic crops to DINOv2 | DINOv2 · ImageNet val cyclic shifts | accuracy (%): 84.1 -> 83.93 | Hurt |
| apply RICS to DINOv2 on Places365 | DINOv2 · Places365 subset | Adversarial robustness (1-NN) 1px: 74.94 -> 96.89 | Improved |
Closest papers by meaning
- TraVeLGAN: Image-To-Image Translation by Transformation Vector Learning (CVPR 2019 · similarity 0.90)
- Convolutional Neural Networks Are Not Invariant to Translation, but They Can Learn to Be (JMLR 2021 · similarity 0.90)
- Benign Examples: Imperceptible Changes Can Enhance Image Translation Performance (AAAI 2020 · similarity 0.90)
- Few-Shot Unsupervised Image-to-Image Translation (ICCV 2019 · similarity 0.90)
- Unsupervised Image-to-Image Translation Networks (NeurIPS 2017 · similarity 0.89)
- Why do deep convolutional networks generalize so poorly to small image transformations? (JMLR 2019 · 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.
