Research map / Architectures and efficiency / CNN design
Designing Affine-Invariant Neural Networks for Photometric Corruption Robustness and Generalization
Mounir Messaoudi, Quentin Rapilly, Sébastien Herbreteau, Anaïs Badoual, Charles Kervrann
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 SEqSI architecture instead of Standard | ResNet-20 · CIFAR-10 shift corruption (no train aug) | accuracy %: 51.1 -> 91.2 | Improved |
| use SEqSI architecture instead of Standard | ResNet-20 · CIFAR-10 affine corruption (no train aug) | accuracy %: 64.1 -> 91.2 | Improved |
| use SEqSI instead of Standard on unseen spatially-varying affine | ResNet-20 · CIFAR-10 (no train aug) | accuracy %: 31.8 -> 72.5 | Improved |
| use SEqSI instead of Standard on high noise | ResNet-20 · CIFAR-10 (no train aug) | accuracy %: 19.0 -> 15.2 | No clear effect |
| use SEqSI instead of Standard on clean images | ResNet-20 · CIFAR-10 original (no train aug) | accuracy %: 91.7 -> 91.2 | Hurt |
| use SEqSI instead of AffEq (less restrictive constraint) | ResNet-20 · CIFAR-10 original (no train aug) | accuracy %: 89.6 -> 91.2 | Improved |
| add affine and non-affine training augmentation (All) to SEqSI (ablation) | SEqSI ResNet-20 · CIFAR-10 spatially-varying affine | accuracy %: 72.5 -> 89.4 | Improved |
| use SEqSI instead of Standard, OOD Denoised domain | 3D ResNet · Cryo-ET Denoised (train WBP) | accuracy %: 22.36 -> 74.53 | Improved |
| use SEqSI instead of Standard, OOD IsoNet Corrected domain | 3D ResNet · Cryo-ET IsoNet Corrected (train WBP) | accuracy %: 15.95 -> 73.21 | Improved |
| use SEqSI instead of Standard, OOD CTF Deconvolved domain | 3D ResNet · Cryo-ET CTF Deconvolved (train WBP) | accuracy %: 48.91 -> 66.51 | Improved |
| use SEqSI instead of Standard, in-distribution WBP | 3D ResNet · Cryo-ET WBP in-distribution | accuracy %: 87.17 -> 85.15 | Hurt |
| use AffEq instead of SEqSI on CTF Deconvolved | 3D ResNet · Cryo-ET CTF Deconvolved | accuracy %: 66.51 -> 79.07 | Mixed |
Closest papers by meaning
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.87)
- A Regularization-Guided Equivariant Approach for Image Restoration (CVPR 2025 · similarity 0.86)
- Improving robustness against common corruptions by covariate shift adaptation (NeurIPS 2020 · similarity 0.86)
- Relaxing Equivariance Constraints with Non-stationary Continuous Filters (NeurIPS 2022 · similarity 0.86)
- The Lie Derivative for Measuring Learned Equivariance (ICLR 2023 · similarity 0.85)
- D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Amplitude and Pixel Spaces (NeurIPS 2026 · similarity 0.84)
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.
