geirhos_conflict_stimuli

  • توضیحات :

محرک‌های تضاد شکل/بافت از «سی‌ان‌ان‌های آموزش‌دیده‌شده توسط ImageNet به سمت بافت سوگیری دارند؛ افزایش سوگیری شکل، دقت و استحکام را بهبود می‌بخشد».

توجه داشته باشید که، اگرچه منبع داده شامل تصاویری با شکل و بافت منطبق است و ما آنها را در اینجا قرار می دهیم، اما برای اکثر ارزیابی ها در مقاله اصلی نادیده گرفته می شوند.

شکاف مثال ها
'test' 1280
  • ساختار ویژگی :
FeaturesDict({
    'file_name': Text(shape=(), dtype=string),
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'shape_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=int64, num_classes=1000)),
    'shape_label': ClassLabel(shape=(), dtype=int64, num_classes=16),
    'texture_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=int64, num_classes=1000)),
    'texture_label': ClassLabel(shape=(), dtype=int64, num_classes=16),
})
  • مستندات ویژگی :
ویژگی کلاس شکل نوع D شرح
FeaturesDict
نام فایل متن رشته
تصویر تصویر (هیچ، هیچ، 3) uint8
shape_imagenet_labels دنباله (ClassLabel) (هیچ یک،) int64
شکل_برچسب ClassLabel int64
texture_imagenet_labels دنباله (ClassLabel) (هیچ یک،) int64
بافت_برچسب ClassLabel int64

تجسم

  • نقل قول :
@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},
}