Research map / Label-efficient and robust learning / Recognition and long-tail
The Origins and Prevalence of Texture Bias in Convolutional Neural Networks
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 |
|---|---|---|---|
| Use center crops instead of random crops (ablation) | ImageNet-trained ResNet-50 · GST | Shape Bias: 19.5% -> 28.4% | Helped |
| Add Gaussian noise (ablation) | ImageNet-trained ResNet-50 · GST | Shape Bias: 19.5% -> 30.7% | Helped |
| Add color distortion, Gaussian blur, and Gaussian noise, then use a minimum crop area of 64% and stronger augmentation (ablation) | ImageNet-trained ResNet-50 · GST | Shape Bias: 19.5% -> 55.2% | Helped |
| Train the stronger-augmentation model for 270 epochs (ablation) | ImageNet-trained ResNet-50 with naturalistic augmentation and larger crops · GST | Shape Bias: 55.2% -> 62.2% | Helped |
| Add naturalistic augmentation and increase minimum crop area, with stronger augmentation (ablation) | ImageNet-trained ResNet-50 · ImageNet-Sketch | top-1 accuracy: 22.4% -> 30.4% | Helped |
| Add naturalistic augmentation and increase minimum crop area, with stronger augmentation (ablation) | ImageNet-trained ResNet-50 · Stylized ImageNet | top-1 accuracy: 7.7% -> 15.1% | Helped |
| Use stronger naturalistic augmentation and larger crops (ablation) | ImageNet-trained ResNet-50 · ImageNet | top-1 accuracy: 76.6% -> 72.0% | Hurt |
| Add rotation augmentation (ablation) | ImageNet-trained ResNet-50 · GST | Shape Bias: 19.5% -> 19.4% | No clear effect |
| Train with rotation classification rather than supervised classification | ResNet-50 representation · GST | Shape Bias: 21.9% -> 32.3% | Helped |
| Use SimCLR augmentation in supervised training instead of the SimCLR objective | ResNet-50 representation · GST | Shape Bias: 37.0% -> 40.4% | Helped |
| Replace convolution with local attention | ResNet-50 · GST | Shape Bias: 23.2% -> 20.2% | No clear effect |
| Increase minimum random-crop area (ablation) | ImageNet-trained ResNet-50 · ImageNet | top-1 accuracy: 76.6% -> 73.5% | Hurt |
Closest papers by meaning
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.95)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.95)
- ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (ICLR 2019 · similarity 0.94)
- BiasBed - Rigorous Texture Bias Evaluation (CVPR 2023 · similarity 0.92)
- Shape-Texture Debiased Neural Network Training (ICLR 2021 · similarity 0.92)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · 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.