geirhos_conflict_stimuli

  • คำอธิบาย :

สิ่งเร้าความขัดแย้งของรูปร่าง/พื้นผิวจาก "ซีเอ็นเอ็นที่ได้รับการฝึกฝนโดย ImageNet มีความเอนเอียงไปทางพื้นผิว การเพิ่มอคติของรูปร่างช่วยเพิ่มความแม่นยำและความทนทาน"

โปรดทราบว่าแม้ว่าแหล่งข้อมูลชุดข้อมูลจะมีรูปภาพที่มีรูปร่างและพื้นผิวตรงกัน และเรารวมไว้ที่นี่ แต่จะถูกละเว้นสำหรับการประเมินส่วนใหญ่ในเอกสารต้นฉบับ

แยก ตัวอย่าง
'test' 1,280
  • โครงสร้างคุณลักษณะ :
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),
})
  • เอกสารคุณสมบัติ :
ลักษณะเฉพาะ ระดับ รูปร่าง Dประเภท คำอธิบาย
คุณสมบัติDict
ชื่อไฟล์ ข้อความ สตริง
ภาพ ภาพ (ไม่มี, ไม่มี, 3) uint8
shape_imagenet_labels ลำดับ (ClassLabel) (ไม่มี,) int64
รูปร่าง_label ป้ายกำกับคลาส int64
texture_imagenet_labels ลำดับ (ClassLabel) (ไม่มี,) int64
texture_label ป้ายกำกับคลาส int64

การสร้างภาพ

  • การอ้างอิง :
@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},
}