geirhos_conflict_stimuli

  • 설명 :

"ImageNet 훈련된 CNN은 텍스처 쪽으로 편향되어 있습니다. 모양 편향이 증가하면 정확도와 견고성이 향상됩니다."의 모양/텍스처 충돌 자극.

데이터 세트 소스에 모양과 질감이 일치하는 이미지가 포함되어 있고 여기에 포함되어 있지만 원본 논문의 대부분의 평가에서는 무시됩니다.

나뉘다
'test' 1,280
  • 기능 구조 :
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),
})
  • 기능 문서 :
특징 수업 모양 D타입 설명
풍모Dict
파일 이름 텍스트
영상 영상 (없음, 없음, 3) uint8
shape_imagenet_labels 시퀀스(클래스 레이블) (없음,) int64
shape_label 클래스 레이블 int64
texture_imagenet_labels 시퀀스(클래스 레이블) (없음,) int64
texture_label 클래스 레이블 int64

심상

  • 인용 :
@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},
}