CITRIS - Causal Representation Learning Datasets

This repository contains the datasets from the paper "CITRIS: Causal Identifiability from Temporal Intervened Sequences" (link) by Phillip Lippe, Sara Magliacane, Sindy Löwe, Yuki M. Asano, Taco Cohen, Efstratios Gavves.

Temporal Causal3DIdent - The Temporal Causal3DIdent dataset is a collection of 3D object shapes, which are observed under varying positions, rotations, lightning, and colors. Overall, we this dataset contains 7 (multidimensional) causal factors. The 7 shapes used are Armadillo, Bunny, Cow, Dragon, Head, Horse, Teapot. We provide two versions of the dataset: one that only contains images of the Teapot, and one that uses all 7 shapes. For more details on the dataset, see our GitHub repository.

Interventional Pong - The Interventional Pong environment is inspired by the game dynamics of Pong, where both paddles follow the policy of moving towards the ball, and the ball has slightly random movements. This dataset considers the 5 causal variables paddle left, paddle right, the ball position, the ball velocity, and the score. For more details on the dataset, see our GitHub repository.

Ball-in-Boxes - The Ball-in-Boxes is a simple dataset for showcasing the concept of the minimal causal variables. The system consists of a ball which randomly moves within a box, but only under an intervention can swap between the two boxes. Thereby, the intervention does not affect the x-position in the box. Thus, one can only discover the box assignment as a causal variable, and not whether the inner x-position also belongs to it. For more details on the dataset, see our GitHub repository.