geirhos_conflict_stimuli

  • opis :

Bodźce konfliktu kształtu / tekstury z „CNN przeszkolonych w ImageNet są ukierunkowane na teksturę; zwiększenie odchylenia kształtu poprawia dokładność i solidność”.

Należy zauważyć, że chociaż źródło zestawu danych zawiera obrazy o pasującym kształcie i teksturze i uwzględniamy je tutaj, są one ignorowane w przypadku większości ocen w oryginalnym artykule.

Rozdzielać Przykłady
'test' 1280
  • Struktura funkcji :
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),
})
  • Dokumentacja funkcji :
Funkcja Klasa Kształt Typ D Opis
FunkcjeDict
Nazwa pliku Tekst strunowy
obraz Obraz (Brak, Brak, 3) uint8
shape_imagenet_labels Sekwencja (etykieta klasy) (Nic,) int64
etykieta_kształtu Etykieta klasy int64
texture_imagenet_labels Sekwencja (etykieta klasy) (Nic,) int64
etykieta_tekstury Etykieta klasy int64

Wyobrażanie sobie

  • Cytat :
@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},
}