ACRE

Abstract Causal REasoning (ACRE) is a dataset for the systematic evaluation of current vision systems in causal induction, i.e., identifying unobservable mechanisms that lead to the observable relations among variables.

Each split of the dataset is structured as follows:

config/
    train.json
    val.json
    test.json
images/
    ACRE_train_00*.png
    ACRE_val_00*.png
    ACRE_test_00*.png
scenes/
    ACRE_train_00*.json
    ACRE_val_00*.json
    ACRE_test_00*.json

Each image file in the images folder has a corresponding scene description file in scenes with the same name (except for the extension).

Each ACRE problem is named after ACRE_{train/val/test}_{6_digit_problem_idx}_{2_digit_panel_idx}