- 説明:
「ImageNet でトレーニングされた CNN はテクスチャに偏っています。形状バイアスを増やすと、精度と堅牢性が向上します。」
データセット ソースには形状とテクスチャが一致する画像が含まれており、ここに含めていますが、元の論文のほとんどの評価では無視されていることに注意してください。
バージョン:
-
1.0.0(デフォルト): リリース ノートはありません。
-
ダウンロードサイズ:
153.96 MiBデータセットのサイズ:
130.44 MiB自動キャッシュ(ドキュメント):
shuffle_files=Falseの場合のみ (テスト)スプリット:
| スプリット | 例 |
|---|---|
'test' | 1,280 |
- 機能構造:
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),
})
- 機能のドキュメント:
| 特徴 | クラス | 形 | Dtype | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| ファイル名 | 文章 | ストリング | ||
| 画像 | 画像 | (なし、なし、3) | uint8 | |
| shape_imagenet_labels | シーケンス(クラスラベル) | (なし、) | int64 | |
| shape_label | クラスラベル | int64 | ||
| texture_imagenet_labels | シーケンス(クラスラベル) | (なし、) | int64 | |
| texture_label | クラスラベル | int64 |
監視キー(
as_superviseddocを参照):('image', 'shape_label')図( tfds.show_examples ):

- 例( tfds.as_dataframe ):
- 引用:
@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},
}