shape3d

3dshapes は、6 つのグラウンド トゥルースに依存しない潜在因子から手続き的に生成された 3D 形状のデータセットです。これらの要因は、床の色壁の色オブジェクトの色スケール形状向きです。

これらの潜在のすべての可能な組み合わせは、N = 480000 の合計画像を生成する 1 回だけ存在します。

潜在因子の値

  • 床色相: [0, 1] で線形に配置された 10 個の値
  • 壁の色相: [0, 1] で線形に配置された 10 個の値
  • オブジェクトの色相: [0, 1] で線形に配置された 10 個の値
  • スケール: [0, 1] で線形に配置された 8 つの値
  • 形状: [0, 1, 2, 3] の 4 つの値
  • 方向: [-30, 30] で直線的に配置された 15 個の値

一度に 1 つの潜在変数を変化させ (方向から開始し、次に形状など)、 images配列に一定の順序でイメージを順次格納しました。因子の対応する値は、 labels配列に同じ順序で格納されます。

スプリット
'train' 480,000
  • 機能構造:
FeaturesDict({
    'image': Image(shape=(64, 64, 3), dtype=uint8),
    'label_floor_hue': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'label_object_hue': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'label_orientation': ClassLabel(shape=(), dtype=int64, num_classes=15),
    'label_scale': ClassLabel(shape=(), dtype=int64, num_classes=8),
    'label_shape': ClassLabel(shape=(), dtype=int64, num_classes=4),
    'label_wall_hue': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'value_floor_hue': float32,
    'value_object_hue': float32,
    'value_orientation': float32,
    'value_scale': float32,
    'value_shape': float32,
    'value_wall_hue': float32,
})
  • 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(64, 64, 3) uint8
label_floor_hueクラスラベルint64
label_object_hueクラスラベルint64
label_orientationクラスラベルint64
label_scaleクラスラベルint64
label_shapeクラスラベルint64
label_wall_hueクラスラベルint64
value_floor_hueテンソルfloat32
value_object_hueテンソルfloat32
value_orientationテンソルfloat32
value_scaleテンソルfloat32
値_形状テンソルfloat32
value_wall_hueテンソルfloat32

視覚化

  • 引用
@misc{3dshapes18,
  title={3D Shapes Dataset},
  author={Burgess, Chris and Kim, Hyunjik},
  howpublished={https://github.com/deepmind/3dshapes-dataset/},
  year={2018}
}