Research map / Learning theory and algorithms / Prediction and decision losses
BiasBed - Rigorous Texture Bias Evaluation
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 texture/style debiasing algorithms rather than ERM on true OOD texture-bias datasets | ResNet-50 image classification · True OOD texture-bias datasets | accuracy | No clear effect |
| Select checkpoints by best validation accuracy instead of the best test epoch | ERM and texture-bias methods · Silhouette | accuracy: 50.8 / 47.3 (-3.5) | Hurt |
| Select checkpoints by best validation accuracy instead of the best test epoch | ERM and texture-bias methods · Edge | accuracy: 32.6 / 22.6 (-10.0) | Hurt |
| Select checkpoints by best validation accuracy instead of the best test epoch | ERM and texture-bias methods · Cue-Conflict | accuracy: 26.4 / 22.2 (-4.2) | Hurt |
| Replace ordinary ERM training with feature augmentation (pAdaIN, SagNet, or InfoDrop) | ResNet-50 · Texture-bias datasets | accuracy | No clear effect |
| Use input augmentation rather than feature augmentation | ResNet-50 · Texture-bias datasets | accuracy | Mixed |
| Use formal Friedman and Nemenyi hypothesis tests instead of ranking methods by average dataset scores | BiasBed algorithm comparisons · True OOD datasets | statistical significance: F=4.98, p=0.00045 | Helped |
| Choose a best-test-epoch oracle checkpoint rather than the best-validation checkpoint | ERM, SagNet, pAdaIN, and InfoDrop · Silhouette, Edge, and Cue-Conflict | accuracy: ERM: -3.5, -10.0, -4.2; SagNet: -4.8, -7.4, -1.3; pAdaIN: -4.2, -6.7, -2.5; InfoDrop: -3.4, -11.7, -3.8 | 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 not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.90)
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.87)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.86)
- Perceptual Inductive Bias Is What You Need Before Contrastive Learning (CVPR 2025 · similarity 0.86)
- Low-Biased General Annotated Dataset Generation (CVPR 2025 · 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.