open_images_v4

  • 説明

Open Images は、画像レベルのラベルとオブジェクト境界ボックスで注釈が付けられた約 900 万個の画像のデータセットです。

V4 のトレーニング セットには、174 万枚の画像上の 600 のオブジェクト クラスに対する 1460 万個のバウンディング ボックスが含まれており、オブジェクト位置アノテーションを備えた既存のデータセットとしては最大のものになります。ボックスは、正確さと一貫性を確保するために、プロのアノテーターによって大部分が手作業で描かれています。画像は非常に多様で、多くの場合、複数のオブジェクト (画像あたり平均 8.4) を含む複雑なシーンが含まれています。さらに、データセットには、数千のクラスにわたる画像レベルのラベルで注釈が付けられます。

スプリット
'test' 125,436
'train' 1,743,042
'validation' 41,620
  • 機能の構造:
FeaturesDict({
   
'bobjects': Sequence({
       
'bbox': BBoxFeature(shape=(4,), dtype=float32),
       
'is_depiction': int8,
       
'is_group_of': int8,
       
'is_inside': int8,
       
'is_occluded': int8,
       
'is_truncated': int8,
       
'label': ClassLabel(shape=(), dtype=int64, num_classes=601),
       
'source': ClassLabel(shape=(), dtype=int64, num_classes=6),
   
}),
   
'image': Image(shape=(None, None, 3), dtype=uint8),
   
'image/filename': Text(shape=(), dtype=string),
   
'objects': Sequence({
       
'confidence': int32,
       
'label': ClassLabel(shape=(), dtype=int64, num_classes=19995),
       
'source': ClassLabel(shape=(), dtype=int64, num_classes=6),
   
}),
   
'objects_trainable': Sequence({
       
'confidence': int32,
       
'label': ClassLabel(shape=(), dtype=int64, num_classes=7186),
       
'source': ClassLabel(shape=(), dtype=int64, num_classes=6),
   
}),
})
  • 機能ドキュメント:
特徴クラスDタイプ説明
特徴辞書
オブジェクト順序
ボオブジェクト/BボックスBBox機能(4,) float32
bobjects/is_depictionテンソルint8
bobjects/is_group_ofテンソルint8
bobjects/is_insideテンソルint8
bobjects/is_occludedテンソルint8
bobjects/is_truncatedテンソルint8
オブジェクト/ラベルクラスラベルint64
オブジェクト/ソースクラスラベルint64
画像画像(なし、なし、3) uint8
画像/ファイル名文章
オブジェクト順序
オブジェクト/信頼テンソルint32
オブジェクト/ラベルクラスラベルint64
オブジェクト/ソースクラスラベルint64
オブジェクト_トレーニング可能順序
オブジェクト_訓練可能/信頼性テンソルint32
オブジェクト_トレーニング可能/ラベルクラスラベルint64
オブジェクト_トレーニング可能/ソースクラスラベルint64
@article{OpenImages,
  author
= {Alina Kuznetsova and
           
Hassan Rom and
           
Neil Alldrin and
           
Jasper Uijlings and
           
Ivan Krasin and
           
Jordi Pont-Tuset and
           
Shahab Kamali and
           
Stefan Popov and
           
Matteo Malloci and
           
Tom Duerig and
           
Vittorio Ferrari},
  title
= {The Open Images Dataset V4: Unified image classification,
           
object detection, and visual relationship detection at scale},
  year
= {2018},
  journal
= {arXiv:1811.00982}
}
@article{OpenImages2,
  author
= {Krasin, Ivan and
           
Duerig, Tom and
           
Alldrin, Neil and
           
Ferrari, Vittorio
           
and Abu-El-Haija, Sami and
           
Kuznetsova, Alina and
           
Rom, Hassan and
           
Uijlings, Jasper and
           
Popov, Stefan and
           
Kamali, Shahab and
           
Malloci, Matteo and
           
Pont-Tuset, Jordi and
           
Veit, Andreas and
           
Belongie, Serge and
           
Gomes, Victor and
           
Gupta, Abhinav and
           
Sun, Chen and
           
Chechik, Gal and
           
Cai, David and
           
Feng, Zheyun and
           
Narayanan, Dhyanesh and
           
Murphy, Kevin},
  title
= {OpenImages: A public dataset for large-scale multi-label and
           multi
-class image classification.},
  journal
= {Dataset available from
             https
://storage.googleapis.com/openimages/web/index.html},
  year
={2017}
}

open_images_v4/original (デフォルト設定)

  • 設定の説明: 元の解像度と品質の画像。

  • データセットのサイズ: 562.42 GiB

  • ( tfds.show_examples ):

視覚化

open_images_v4/300k

  • 構成の説明: 画像には、72 JPEG 品質で約 300,000 ピクセルがあります。

  • データセットのサイズ: 81.92 GiB

  • ( tfds.show_examples ):

視覚化

open_images_v4/200k

  • 構成の説明: 画像には、72 JPEG 品質で約 200,000 ピクセルがあります。

  • データセットのサイズ: 60.70 GiB

  • ( tfds.show_examples ):

視覚化