The images were created such that they have an unclear ground truth,
i.e., such that they are similar to multiple - but not all - of the datasets classes.
Robust and uncertainty-aware models should be able to detect and flag these ambiguous images.
As such, the dataset should be merged / mixed with the original dataset and we
provide such 'mixed' splits for convenience. Please refer to the dataset card for details.
