- Mô tả:
Các kích thích xung đột về hình dạng / kết cấu từ "Các CNN được đào tạo bởi ImageNet thiên về kết cấu; sự thiên vị về hình dạng tăng lên sẽ cải thiện độ chính xác và độ chắc chắn."
Lưu ý rằng, mặc dù nguồn tập dữ liệu có chứa các hình ảnh có hình dạng và kết cấu phù hợp và chúng tôi đưa chúng vào đây, chúng sẽ bị bỏ qua đối với hầu hết các đánh giá trong bài báo gốc.
Trang chủ: https://github.com/rgeirhos/texture-vs-shape
Source code:
tfds.image_classification.GeirhosConflictStimuli
phiên bản:
-
1.0.0
(mặc định): Không có ghi chú phát hành.
-
Dung lượng tải về:
153.96 MiB
Dataset kích thước:
130.44 MiB
Tự động lưu trữ ( tài liệu ): Chỉ khi
shuffle_files=False
(thử nghiệm)tách:
Tách ra | Các ví dụ |
---|---|
'test' | 1.280 |
- Các tính năng:
FeaturesDict({
'file_name': Text(shape=(), dtype=tf.string),
'image': Image(shape=(None, None, 3), dtype=tf.uint8),
'shape_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=tf.int64, num_classes=1000)),
'shape_label': ClassLabel(shape=(), dtype=tf.int64, num_classes=16),
'texture_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=tf.int64, num_classes=1000)),
'texture_label': ClassLabel(shape=(), dtype=tf.int64, num_classes=16),
})
Phím giám sát (Xem
as_supervised
doc ):('image', 'shape_label')
Hình ( tfds.show_examples ):
- Ví dụ ( tfds.as_dataframe ):
- Trích dẫn:
@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},
}