food101

  • 説明:

このデータセットは、101,000 枚の画像を含む 101 の食品カテゴリで構成されています。クラスごとに、手動でレビューされた 250 のテスト画像と 750 のトレーニング画像が提供されます。トレーニング イメージは意図的にクリーニングされていないため、まだある程度のノイズが含まれています。これは主に強烈な色と、時には間違ったラベルの形で発生します.すべての画像は、最大辺の長さが 512 ピクセルになるように再スケーリングされています。

スプリット
'train' 75,750
'validation' 25,250
  • 機能構造:
FeaturesDict({
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=101),
})
  • 機能のドキュメント:
特徴クラスDtype説明
特徴辞書
画像画像(なし、なし、3) uint8
ラベルクラスラベルint64

視覚化

  • 引用
@inproceedings{bossard14,
  title = {Food-101 -- Mining Discriminative Components with Random Forests},
  author = {Bossard, Lukas and Guillaumin, Matthieu and Van Gool, Luc},
  booktitle = {European Conference on Computer Vision},
  year = {2014}
}