open_images_challenge2019_detection

Open Images is a collaborative release of ~9 million images annotated with image-level labels, object bounding boxes, object segmentation masks, and visual relationships. This uniquely large and diverse dataset is designed to spur state of the art advances in analyzing and understanding images.

This contains the data from thee Object Detection track of the competition. The goal in this track is to predict a tight bounding box around all object instances of 500 classes.

The images are annotated with positive image-level labels, indicating certain object classes are present, and with negative image-level labels, indicating certain classes are absent. In the competition, all other unannotated classes are excluded from evaluation in that image. For each positive image-level label in an image, every instance of that object class in the image was annotated.

Split Examples
'test' 99,999
'train' 1,743,042
'validation' 41,620
  • Feature structure:
FeaturesDict({
    'bobjects': Sequence({
        'bbox': BBoxFeature(shape=(4,), dtype=tf.float32),
        'is_group_of': tf.bool,
        'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=500),
    }),
    'id': Text(shape=(), dtype=tf.string),
    'image': Image(shape=(None, None, 3), dtype=tf.uint8),
    'objects': Sequence({
        'confidence': tf.float32,
        'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=500),
        'source': Text(shape=(), dtype=tf.string),
    }),
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
bobjects Sequence
bobjects/bbox BBoxFeature (4,) tf.float32
bobjects/is_group_of Tensor tf.bool
bobjects/label ClassLabel tf.int64
id Text tf.string
image Image (None, None, 3) tf.uint8
objects Sequence
objects/confidence Tensor tf.float32
objects/label ClassLabel tf.int64
objects/source Text tf.string

open_images_challenge2019_detection/200k (default config)

  • Config description: Images have at most 200,000 pixels, at 72 JPEG quality.

  • Dataset size: 59.40 GiB

open_images_challenge2019_detection/300k

  • Config description: Images have at most 300,000 pixels, at 72 JPEG quality.

  • Dataset size: 80.44 GiB