---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: file_id
    dtype: string
  - name: shape_cue
    dtype: string
  - name: texture_cue
    dtype: string
  splits:
  - name: test
    num_bytes: 136418289
    num_examples: 1280
  download_size: 136531921
  dataset_size: 136418289
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
task_categories:
- image-classification
pretty_name: texture-shape cue conflict
size_categories:
- 1K<n<10K
arxiv: 2509.25339
license: cc-by-4.0
---

This dataset contains the stimuli for [ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness](https://openreview.net/forum?id=Bygh9j09KX) by Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. 
The stimuli allow testing of the texture/shape bias in an observer model (human or artificial) by containing two conflicting cues per image (shape and texture). The images generated using iterative style transfer (Gatys et al., 2016) between an image of the texture data set (as style) and an image from the original data set (as content).
There is a total of 1280 cue conflict images (80 per category) and 16 classes.

## 📚 Citation

```latex
@inproceedings{geirhos2019imagenettrained,
  title={ImageNet-trained {CNN}s are biased towards texture; increasing shape bias improves accuracy and robustness.},
  author={Robert Geirhos and Patricia Rubisch and Claudio Michaelis and Matthias Bethge and Felix A. Wichmann and Wieland Brendel},
  booktitle={International Conference on Learning Representations},
  year={2019},
  url={https://openreview.net/forum?id=Bygh9j09KX},
}
```