Multi-Object Datasets

This repository contains datasets for multi-object representation learning, used in developing scene decomposition methods like MONet [1], IODINE [2], and SIMONe [3]. The datasets we provide are:

1. Multi-dSprites
2. Objects Room
3. CLEVR (with masks)
4. Tetrominoes
5. CATER (with masks)

The datasets consist of multi-object scenes. Each image or video is accompanied by ground-truth segmentation masks for all objects in the scene. For some datasets (excluding Objects Room and CATER), we also provide per-object generative factors to facilitate representation learning. The generative factors include all necessary and sufficient features (size, color, position, etc.) to describe and render the objects present in a scene.

Lastly, the segmentation_metrics module contains a TensorFlow implementation of the adjusted Rand index [4], which can be used to compare inferred object segmentations with ground-truth segmentation masks. All code has been tested to work with TensorFlow r1.14.

Bibtex

If you use one of these datasets in your work, please cite it as follows:

@misc{multiobjectdatasets19,
  title={Multi-Object Datasets},
  author={Kabra, Rishabh and Burgess, Chris and Matthey, Loic and
          Kaufman, Raphael Lopez and Greff, Klaus and Reynolds, Malcolm and
          Lerchner, Alexander},
  howpublished={https://github.com/deepmind/multi-object-datasets/},
  year={2019}
}

Descriptions

Multi-dSprites

This is a dataset based on dSprites. Each image consists of multiple oval, heart, or square-shaped sprites (with some occlusions) set against a uniformly colored background.

We're releasing three versions of this dataset containing 1M datapoints each:

1.1 Binarized: each image has 2-3 white sprites on a black background.

1.2 Colored sprites on grayscale: each scene has 2-5 randomly colored HSV sprites on a randomly sampled grayscale background.

1.3 Colored sprites and background: each scene has 1-4 sprites. All colors are randomly sampled RGB values.

Each datapoint contains an image, a number of background and object masks, and the following ground-truth features per object: x and y positions, shape, color (rgb values), orientation, and scale. Lastly, visibility is a binary feature indicating which objects are not null.

Objects Room

This dataset is based on the MuJoCo environment used by the Generative Query Network [5] and is a multi-object extension of the 3d-shapes dataset. The training set contains 1M scenes with up to three objects. We also provide ~1K test examples for the following variants:

2.1 Empty room: scenes consist of the sky, walls, and floor only.

2.2 Six objects: exactly 6 objects are visible in each image.

2.3 Identical color: 4-6 objects are placed in the room and have an identical, randomly sampled color.

Datapoints consist of an image and fixed number of masks. The first four masks correspond to the sky, floor, and two halves of the wall respectively. The remaining masks correspond to the foreground objects.

CLEVR (with masks)

We adapted the open-source script provided by Johnson et al. to produce ground-truth segmentation masks for CLEVR [6] scenes. These were generated afresh, so images in this dataset are not identical to those in the original CLEVR dataset. We ignore the original question-answering task.

The images and masks in the dataset are of size 320x240. We also provide all ground-truth factors included in the original dataset (namely x, y, and z position, pixel_coords, and rotation, which are real-valued; plus size, material, shape, and color, which are encoded as integers) along with a visibility vector to indicate which objects are not null.

Tetrominoes

This is a dataset of Tetris-like shapes (aka tetrominoes). Each 35x35 image contains three tetrominoes, sampled from 17 unique shapes/orientations. Each tetromino has one of six possible colors (red, green, blue, yellow, magenta, cyan). We provide x and y position, shape, and color (integer-coded) as ground-truth features. Datapoints also include a visibility vector.

CATER (with masks)

We adapted the open-source script provided by Girdhar et al. to produce ground-truth segmentation masks for CATER [7] videos. We use identical settings as the max2action_cameramotion version of the dataset, containing a moving camera and up to two moving objects at any time. We ignore the original tasks.

The videos and masks we provide are of size 64x64, obtained by taking a central crop and downscaling the original 320x240 images. Each video contains 33 frames. For each frame we also provide a 4x4 camera_matrix containing the orientation and position of the camera, and object_positions containing the 3D allocentric positions of all objects in the scene.

Note that each split (train and test) of this dataset is sharded across 100 TFRecord files. To load either split fully, pass all corresponding filenames into the dataset loader.

Download

The datasets can be downloaded from Google Cloud Storage. Each dataset is a single TFRecords file. To download a particular dataset, use the web interface, or run wget with the appropriate filename as follows:

  wget https://storage.googleapis.com/multi-object-datasets/multi_dsprites/multi_dsprites_colored_on_colored.tfrecords

To download all datasets, you'll need the gsutil tool, which comes with the Google Cloud SDK. Simply run:

  gsutil cp -r gs://multi-object-datasets .

The approximate download sizes are:

1. Multi-dSprites: between 500 MB and 1 GB.
2. Objects Room: the training set is 7 GB. The test sets are 6-8 MB.
3. CLEVR (with masks): 10.5 GB.
4. Tetrominoes: 300 MB.
5. CATER (with masks): the training set is 8 GB. The test set is 4 GB.

Usage

After downloading the dataset files, you can read them as tf.data.Dataset instances with the readers provided. The example below shows how to read the colored-sprites-and-background version of Multi-dSprites:

  from multi_object_datasets import multi_dsprites
  import tensorflow as tf

  tf_records_path = 'path/to/multi_dsprites_colored_on_colored.tfrecords'
  batch_size = 32

  dataset = multi_dsprites.dataset(tf_records_path, 'colored_on_colored')
  batched_dataset = dataset.batch(batch_size)  # optional batching
  iterator = batched_dataset.make_one_shot_iterator()
  data = iterator.get_next()

  with tf.train.SingularMonitoredSession() as sess:
    d = sess.run(data)

All dataset readers return images and segmentation masks in the following canonical format (assuming the dataset is batched as above):

• 'image': Tensor of shape [batch_size, height, width, channels] and type uint8. For video datasets, the shape is [batch_size, sequence_length, height, width, channels].

• 'mask': Tensor of shape [batch_size, max_num_entities, height, width, channels] and type uint8. For video datasets, the shape is [batch_size, sequence_length, max_num_entities, height, width, channels]. The tensor takes on values of 255 or 0, denoting whether a pixel belongs to a particular entity or not.

You can compare predicted object segmentation masks with the ground-truth masks using segmentation_metrics.adjusted_rand_index as below:

  max_num_entities = multi_dsprites.MAX_NUM_ENTITIES['colored_on_colored']
  # Ground-truth segmentation masks are always returned in the canonical
  # [batch_size, max_num_entities, height, width, channels] format. To use these
  # as an input for `segmentation_metrics.adjusted_rand_index`, we need them in
  # the [batch_size, n_points, n_true_groups] format,
  # where n_true_groups == max_num_entities. We implement this reshape below.
  # Note that 'oh' denotes 'one-hot'.
  desired_shape = [batch_size,
                   multi_dsprites.IMAGE_SIZE[0] * multi_dsprites.IMAGE_SIZE[1],
                   max_num_entities]
  true_groups_oh = tf.transpose(data['mask'], [0, 2, 3, 4, 1])
  true_groups_oh = tf.reshape(true_groups_oh, desired_shape)

  random_prediction = tf.random_uniform(desired_shape[:-1],
                                        minval=0, maxval=max_num_entities,
                                        dtype=tf.int32)
  random_prediction_oh = tf.one_hot(random_prediction, depth=max_num_entities)

  ari = segmentation_metrics.adjusted_rand_index(true_groups_oh,
                                                 random_prediction_oh)

To exclude all background pixels from the ARI score (as in [2]), you can compute it as follows instead. This assumes the first true group contains all background pixels:

  ari_nobg = segmentation_metrics.adjusted_rand_index(true_groups_oh[..., 1:],
                                                      random_prediction_oh)

References

[1] Burgess, C. P., Matthey, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., & Lerchner, A. (2019). Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390.

[2] Greff, K., Kaufman, R. L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthey, L., Botvinick, M., & Lerchner, A. (2019). Multi-Object Representation Learning with Iterative Variational Inference. Proceedings of the 36th International Conference on Machine Learning, in PMLR 97:2424-2433.

[3] Kabra, R., Zoran, D., Erdogan, G., Matthey, L., Creswell, A., Botvinick, M., Lerchner, A., & Burgess, C. P. (2021). SIMONe: View-Invariant, Temporally-Abstracted Object Representations via Unsupervised Video Decomposition. Advances in Neural Information Processing Systems.

[4] Rand, W. M. (1971). Objective criteria for the evaluation of clustering methods. Journal of the American Statistical association, 66(336), 846-850.

[5] Eslami, S., Rezende, D. J., Besse, F., Viola, F., Morcos, A., Garnelo, M., Ruderman, A., Rusu, A., Danihelka, I., Gregor, K., Reichert, D., Buesing, L., Weber, T., Vinyals, O., Rosenbaum, D., Rabinowitz, N., King, H., Hillier, C., Botvinick, M., Wierstra, D., Kavukcuoglu, K., & Hassabis, D. (2018). Neural scene representation and rendering. Science, 360(6394), 1204-1210.

[6] Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., & Girshick, R. (2017). Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 2901-2910).

[7] Girdhar, R., & Ramanan, D. (2019, September). CATER: A diagnostic dataset for Compositional Actions & TEmporal Reasoning. In International Conference on Learning Representations.

Disclaimers

This is not an official Google product.