Research map / Adversarial robustness and security / Attacks on text and graphs
Shape-Texture Debiased Neural Network Training
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 |
|---|---|---|---|
| train on cue-conflict images with soft labels supervised by both shape and texture | ResNet-50 · ImageNet | top-1 accuracy: 76.4 -> 76.9 (+0.5) | Helped |
| train on cue-conflict images with soft labels supervised by both shape and texture | ResNet-152 · ImageNet-A | accuracy: 7.4 -> 12.6 (+5.2%) | Helped |
| train on cue-conflict images with soft labels supervised by both shape and texture | ResNet-152 · ImageNet-C | mCE: 67.2 -> 58.9 (-8.3%) | Helped |
| train on cue-conflict images with soft labels supervised by both shape and texture | ResNet-152 · Stylized-ImageNet | accuracy: 11.3 -> 22.4 (+11.1%) | Helped |
| train on cue-conflict images with soft labels supervised by both shape and texture | ResNet-152 · ImageNet FGSM | accuracy: 25.2 -> 39.6 (+14.4%) | Helped |
| double the vanilla training schedule | ResNet-101 · ImageNet | top-1 accuracy: 78.0 -> 78.0 (+0.0) | No clear effect |
| assign cue-conflict image labels from shape only (ablation) | ResNet-50 · ImageNet | top-1 accuracy: 76.4 -> 76.2 (-0.2) | Hurt |
| assign cue-conflict image labels from texture only (ablation) | ResNet-50 · ImageNet | top-1 accuracy: 76.4 -> 75.3 (-1.1) | Hurt |
| set shape-texture coefficient γ to 0.7 (ablation) | shape-texture debiased ResNet-50 · ImageNet | top-1 accuracy: 77.0% | Helped |
| vary γ within [0.5, 0.9] (ablation) | shape-texture debiased ResNet-50 · ImageNet | top-1 accuracy | Helped |
| combine shape-texture debiased training with CutMix | ResNeXt-101 + CutMix · ImageNet | top-1 accuracy: 80.5 -> 81.2 (+0.7%) | Helped |
| train with shape-texture debiased supervision (ablation) | ResNet-50 · ImageNet-Sketch | accuracy: 23.8 -> 28.4 | Helped |
Closest papers by meaning
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (NeurIPS 2020 · similarity 0.92)
- ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (ICLR 2019 · similarity 0.91)
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.91)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.90)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.85)
- Harnessing Edge Information for Improved Robustness in Vision Transformers (AAAI 2024 · 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.