Skip to content
No description, website, or topics provided.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
LICENSE.txt Create LICENSE.txt Jul 25, 2019 Update Jul 25, 2019 Python script to read the images from a directory and make a numpy array Jun 7, 2019

MPI3D Disentanglement datasets

MPI3D datasets have been introduced as a part of NeurIPS 2019 Disentanglement Competition. There are three different datasets:

  1. Simplistic rendered images (mpi3d_toy).
  2. Realistic rendered images (mpi3d_realistic).
  3. Real world images (mpi3d_real).

Each dataset consists of 460800 images corresponding to all possible combinations of the following factors of variation:

Factors Possible Values
object_color green=0, red=1, blue=2, brown=3
object_shape cone=0, cube=1, hexagonal prism=2, sphere=3
object_size small=0, large=1
camera_height top=0, center=1, bottom=2
background_color purple=0, sea green=1, salmon=2
horizontal_axis 0,...,39
vertical_axis 0,...,39

Each image has as filename padded_index.png where
index = object_color * 115200 + object_shape * 28800 + object_size * 14400 + camera_height * 4800 + background_color * 1600 + horizontal_axis * 40 + vertical_axis
padded_index = index padded with zeros such that it has 6 digits.

If you use python, this means that once the data is loaded into a numpy array you can use array.reshape([4,4,2,3,3,40,40]) to obtain an array where each dimension corresponds to a factor. Size of images for the simplistic rendered dataset are 64x64.

For more details on the dataset please consult For loading the dataset you may make use of the python scripts in this repository. If you use this dataset then kindly cite us.

  title={On the Transfer of Inductive Bias from Simulation to the Real World: a New Disentanglement Dataset},
  author={Gondal, Muhammad Waleed and W{\"u}thrich, Manuel and Miladinovi{\'c}, {\DJ}or{\dj}e and Locatello, Francesco and Breidt, Martin and Volchkov, Valentin and Akpo, Joel and Bachem, Olivier and Sch{\"o}lkopf, Bernhard and Bauer, Stefan},
  journal={arXiv preprint arXiv:1906.03292},

Links to datasets

simplistic rendered:
realistic rendered: not yet published
real images: not yet published


This work is licensed under a Creative Commons Attribution 4.0 International License (

You can’t perform that action at this time.