CLEVR-Hans

The CLEVR-Hans data set is a novel confounded visual scene data set, which captures complex compositions of different objects. This data set consists of CLEVR images divided into several classes.

The membership of a class is based on combinations of objects’ attributes and relations. Additionally, certain classes within the data set are confounded. Thus, within the data set, consisting of train, validation, and test splits, all train, and validation images of confounded classes will be confounded with a specific attribute or combination of attributes.

Each class is represented by 3000 training images, 750 validation images, and 750 test images. The training, validation, and test set splits contain 9000, 2250, and 2250 samples, respectively, for CLEVR-Hans3 and 21000, 5250, and 5250 samples for CLEVR-Hans7. The class distribution is balanced for all data splits.

For CLEVR-Hans classes for which class rules contain more than three objects, the number of objects to be placed per scene was randomly chosen between the minimal required number of objects for that class and ten, rather than between three and ten, as in the original CLEVR data set.

Finally, the images were created such that the exact combinations of the class rules did not occur in images of other classes. It is possible that a subset of objects from one class rule occur in an image of another class. However, it is not possible that more than one complete class rule is contained in an image.

Source: CLEVR-Hans Image Source: CLEVR-Hans