geirhos_conflict_stimuli

  • বর্ণনা :

আকৃতি/টেক্সচারের দ্বন্দ্বের উদ্দীপনা "ইমেজনেট-প্রশিক্ষিত সিএনএনগুলি টেক্সচারের প্রতি পক্ষপাতদুষ্ট; আকারের পক্ষপাত বৃদ্ধি সঠিকতা এবং দৃঢ়তা উন্নত করে।"

মনে রাখবেন, যদিও ডেটাসেট উৎসে আকৃতি এবং টেক্সচারের সাথে মিলে যাওয়া ছবি রয়েছে এবং আমরা সেগুলিকে এখানে অন্তর্ভুক্ত করি, তবে মূল কাগজে বেশিরভাগ মূল্যায়নের জন্য সেগুলি উপেক্ষা করা হয়।

বিভক্ত উদাহরণ
'test' 1,280
  • বৈশিষ্ট্য গঠন :
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),
})
  • বৈশিষ্ট্য ডকুমেন্টেশন :
বৈশিষ্ট্য ক্লাস আকৃতি ডিটাইপ বর্ণনা
ফিচারসডিক্ট
ফাইল_নাম পাঠ্য স্ট্রিং
ইমেজ ছবি (কোনটিই নয়, 3) uint8
আকৃতি_ইমেজনেট_লেবেল সিকোয়েন্স (ক্লাসলেবেল) (কোনটিই নয়,) int64
আকৃতি_লেবেল ক্লাসলেবেল int64
টেক্সচার_ইমেজেনেট_লেবেল সিকোয়েন্স (ক্লাসলেবেল) (কোনটিই নয়,) int64
টেক্সচার_লেবেল ক্লাসলেবেল int64

ভিজ্যুয়ালাইজেশন

  • উদ্ধৃতি :
@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},
}