geirhos_conflict_stimuli

  • الوصف :

محفزات تعارض الشكل / الملمس من "شبكات CNN المدربة على ImageNet منحازة نحو النسيج ؛ زيادة التحيز في الشكل يحسن الدقة والمتانة."

لاحظ أنه على الرغم من أن مصدر مجموعة البيانات يحتوي على صور ذات شكل ونسيج متطابقين ونقوم بتضمينها هنا ، إلا أنه يتم تجاهلها في معظم التقييمات في الورقة الأصلية.

انشق، مزق أمثلة
'test' 1،280
  • هيكل الميزة :
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),
})
  • وثائق الميزة :
ميزة فصل شكل نوع وصف
الميزات
اسم الملف نص سلسلة
صورة صورة (لا شيء ، لا شيء ، 3) uint8
الشكل_التصوير_التسميات تسلسل (ClassLabel) (لا أحد،) int64
شكل_تصنيف ClassLabel int64
text_imagenet_labels تسلسل (ClassLabel) (لا أحد،) int64
نسيج_تصنيف ClassLabel int64

التصور

  • الاقتباس :
@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},
}