Research map / Label-efficient and robust learning / Recognition and long-tail
Do Image Classifiers Generalize Across Time?
Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, Ludwig Schmidt
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 |
|---|---|---|---|
| add noise/corruption augmentation to ILSVRC training (ablation) | ResNet-50 · ImageNet-Vid-Robust | accuracy drop: 15.0 -> 15.6 | No clear effect |
| add noise/corruption augmentation to ILSVRC training (ablation) | ResNet-50 · YTBB-Robust | accuracy drop: 13.2 -> 16.6 | Hurt |
| use an l_inf adversarially robust model (ablation) | ResNeXt-101 · ImageNet-Vid-Robust | accuracy drop: 15.0 -> 12.4 | Mixed |
| use an l_inf adversarially robust model (ablation) | ResNeXt-101 · YTBB-Robust | accuracy drop: 13.2 -> 10.4 | Mixed |
| fine-tune on ImageNet-Vid frames (ablation) | ResNet-50 · ImageNet-Vid-Robust | accuracy drop: 15.0 -> 15.1 | No clear effect |
| fine-tune on Youtube-BB frames (ablation) | ResNet-50 · YTBB-Robust | accuracy drop: 13.2 -> 9.4 | Improved |
| fine-tune on ImageNet-Vid frames (ablation) | ResNet-50 · ImageNet-Vid-Robust | perturbed accuracy: 52.5 -> 65.7 | Improved |
| train a Faster R-CNN detector instead of a classifier (ablation) | ResNet-50 fine-tuned on ImageNet-Vid · ImageNet-Vid-Robust | original accuracy: 80.8 -> 77.6 | Hurt |
| add bounding-box supervision (detection fine-tuning) (ablation) | ResNet-50 · ImageNet-Vid-Robust | accuracy drop: 15.1 -> 12.3 | No clear effect |
| switch from detection to localization only (ablation) | Faster R-CNN R50 · ImageNet-Vid-Robust | mAP drop: 14.0 -> 12.4 | No clear effect |
| add human review of frame similarity and labels (ablation) | ResNet-152 FT · ImageNet-Vid-Robust | accuracy drop: 16.2 -> 14.4 | Improved |
| add human review of frame similarity and labels (ablation) | ResNet-152 FT · YTBB-Robust | accuracy drop: 10.0 -> 8.9 | Improved |
Closest papers by meaning
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019 · similarity 0.87)
- Is Robustness the Cost of Accuracy? -- A Comprehensive Study on the Robustness of 18 Deep Image Classification Models (ECCV 2018 · similarity 0.86)
- I Am Going MAD: Maximum Discrepancy Competition for Comparing Classifiers Adaptively (ICLR 2020 · similarity 0.86)
- Generalisation in humans and deep neural networks (NeurIPS 2018 · similarity 0.86)
- Do ImageNet Classifiers Generalize to ImageNet? (ICML 2019 · similarity 0.86)
- A Comprehensive Study on Robustness of Image Classification Models: Benchmarking and Rethinking (IJCV 2024 · similarity 0.85)
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.
