Research map / 3D vision / 3D generation
Finding Differences Between Transformers and ConvNets Using Counterfactual Simulation Testing
Nataniel Ruiz, Sarah Adel Bargal, Cihang Xie, Kate Saenko, Stan Sclaroff
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 ConvNext-T instead of Swin-T | ConvNext vs Swin · NVD | top-1 accuracy: 21.1 -> 24.7 | Improved |
| use ConvNext-S instead of Swin-S despite lower ImageNet accuracy | ConvNext vs Swin · NVD | top-1 accuracy: 22.1 -> 27.7 | Improved |
| use ConvNext-L-22k instead of Swin-L-22k | ConvNext vs Swin · NVD | top-5 accuracy: 61.4 -> 67.6 | Improved |
| use ConvNext-B instead of Swin-B | ConvNext vs Swin · NVD | top-1 accuracy: 22.7 -> 23.5 | Improved |
| pretrain on ImageNet-22k (Base size, ConvNext) | ConvNext-B · NVD | top-1 accuracy: 23.5 -> 34.0 | Improved |
| pretrain on ImageNet-22k (Base size, Swin) | Swin-B · NVD | top-1 accuracy: 22.7 -> 31.5 | Improved |
| increase network scale Tiny to Small | ConvNext · NVD | top-1 accuracy: 24.7 -> 27.7 | Improved |
| increase network scale Small to Base (ImageNet-1k) | ConvNext · NVD | top-1 accuracy: 27.7 -> 23.5 | Hurt |
| use Swin instead of ConvNext under random patch drop | Swin vs ConvNext · ImageNet val patch drop | accuracy | Mixed |
| use Swin instead of ConvNext under maximal object occlusion | Swin vs ConvNext · NVD occlusion | proportion of correct conserved predictions | Improved |
| use ConvNext instead of Swin on small object scale | ConvNext vs Swin · NVD scale | proportion of correct conserved predictions | Improved |
| use ConvNext instead of Swin on top views | ConvNext vs Swin · NVD top-to-frontal view | proportion of correct conserved predictions: ConvNext-L-22k around 73% | Improved |
Closest papers by meaning
- Progress and Limitations of Deep Networks to Recognize Objects in Unusual Poses (AAAI 2023 · similarity 0.87)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · similarity 0.87)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.84)
- Explicitly Modeling Subcortical Vision with a Neuro-Inspired Front-End Improves CNN Robustness (NeurIPS 2025 · similarity 0.84)
- Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (CVPR 2026 · similarity 0.84)
- Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference (CVPR 2025 · 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.
