Research map / Label-efficient and robust learning / Recognition and long-tail
Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition
Xilin He, Qinliang Lin, Cheng Luo, Weicheng Xie, Siyang Song, Feng Liu, Linlin Shen
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 |
|---|---|---|---|
| add all three modules (EMSE, TSD, TSG) to the Canny+Pix2Pix baseline (ablation) | SDbOA · CIFAR-10 PGD | robust accuracy (%): 46.78 -> 50.91 | Improved |
| add all three modules to the baseline (ablation) | SDbOA · CIFAR-10 C&W | robust accuracy (%): 45.48 -> 56.06 | Improved |
| add all three modules to the baseline (ablation) | SDbOA · CIFAR-10 | clean accuracy (%): 74.32 -> 83.27 | Improved |
| increase TPS deformation intensity lambda beyond 0.1 (ablation) | SDbOA · shape-bias metrics | sbGE and sbIS | Hurt |
| use SDbOA augmentation instead of standard training on ImageNet | ResNet-50 · ImageNet cue-conflict | sbGE (%): 21.39 -> 71.28 | Improved |
| use SDbOA instead of PixMix on common corruptions | WideResNet-40-4 · CIFAR-10-C | mCE (%): 9.5 vs 18.8 | Hurt |
| use SDbOA instead of AugMax on adversarial error | WideResNet-40-4 · CIFAR-10 PGD-20 | misclassification rate (%): 39.1 -> 27.7 | Improved |
| use SDbOA instead of generative adversarial training (FDA, PORT) with lower runtime | WideResNet-34-10 · CIFAR-10 AutoAttack | robust accuracy (%): 60.60 -> 73.36 | Improved |
| use SDbOA on clean data instead of vanilla training | ResNet18 · CIFAR-10 | clean accuracy (%): 90.70 -> 83.27 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- AugMix — AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Edges to Shapes to Concepts: Adversarial Augmentation for Robust Vision (CVPR 2023 · similarity 0.90)
- Geometric and Textural Augmentation for Domain Gap Reduction (CVPR 2022 · similarity 0.88)
- Shape-Texture Debiased Neural Network Training (ICLR 2021 · similarity 0.87)
- ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (ICLR 2019 · similarity 0.87)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.87)
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.87)
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.
