geirhos_conflict_stimuli

  • Description:

Shape/texture conflict stimuli from "ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness."

Note that, although the dataset source contains images with matching shape and texture and we include them here, they are ignored for most evaluations in the original paper.

Split Examples
'test' 1,280
  • Features:
FeaturesDict({
    'file_name': Text(shape=(), dtype=tf.string),
    'image': Image(shape=(None, None, 3), dtype=tf.uint8),
    'shape_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=tf.int64, num_classes=1000)),
    'shape_label': ClassLabel(shape=(), dtype=tf.int64, num_classes=16),
    'texture_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=tf.int64, num_classes=1000)),
    'texture_label': ClassLabel(shape=(), dtype=tf.int64, num_classes=16),
})
@inproceedings{
  geirhos2018imagenettrained,
  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},
}

Visualization