Datasets used to train Generative Query Networks (GQNs) in the ‘Neural Scene Representation and Rendering’ paper.

The following version of the datasets are available:

• rooms_ring_camera. Scenes of a variable number of random objects captured in a square room of size 7x7 units. Wall textures, floor textures as well as the shapes of the objects are randomly chosen within a fixed pool of discrete options. There are 5 possible wall textures (red, green, cerise, orange, yellow), 3 possible floor textures (yellow, white, blue) and 7 possible object shapes (box, sphere, cylinder, capsule, cone, icosahedron and triangle). Each scene contains 1, 2 or 3 objects. In this simplified version of the dataset, the camera only moves on a fixed ring and always faces the center of the room. This is the ‘easiest’ version of the dataset, use version for fast training.

• rooms_free_camera_no_object_rotations. As in rooom_ring_camera, except the camera moves freely. However the objects themselves do not rotate around their axes, which makes the modeling task somewhat easier. This version is ‘medium’ difficulty.

• rooms_free_camera_with_object_rotations. As in rooms_free_camera_no_object_rotations, the camera moves freely, however objects can rotate around their vertical axes across scenes. This is the ‘hardest’ version of the dataset.

• jaco. a reproduction of the robotic Jaco arm is placed in the middle of the room along with one spherical target object. The arm has nine joints. As above, the appearance of the room is modified for each episode by randomly choosing a different texture for the walls and floor from a fixed pool of options. In addition, we modify both colour and position of the target randomly. Finally, the joint angles of the arm are also initialised at random within a range of physically sensible positions.

• shepard_metzler_5_parts. Each object is composed of 7 randomly coloured cubes that are positioned by a self-avoiding random walk in 3D grid. As above, the camera is parametrised by its position, yaw and pitch, however it is constrained to only move around the object at a fixed distance from its centre. This is the ‘easy’ version of the dataset, where each object is composed of only 5 parts.

• shepard_metzler_7_parts. This is the ‘hard’ version of the above dataset, where each object is composed of 7 parts.

• mazes. Random mazes that were created using an OpenGL-based DeepMind Lab game engine (Beattie et al., 2016). Each maze is constructed out of an underlying 7 by 7 grid, with walls falling on the boundaries of the grid locations. However, the agent can be positioned at any continuous position in the maze. The mazes contain 1 or 2 rooms, with multiple connecting corridors. The walls and floor textures of each maze are determined by random uniform sampling from a predefined set of textures.

Usage example

To select what dataset to load, instantiate a reader passing the correct version argument. Note that the constructor will set up all the queues used by the reader. To get tensors call read on the data reader passing in the desired batch size.

  import tensorflow as tf

  root_path = 'path/to/datasets/root/folder'
  data_reader = DataReader(dataset='jaco', context_size=5, root=root_path)
  data = data_reader.read(batch_size=12)

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

Download

Raw data files referred to in this document are available to download here. To download the datasets you can use the gsutil cp command; see also the gsutil installation instructions.

Notes

This is not an official Google product.