So that copies like https://github.com/facebookresearch/detr/blob/master/datasets/coco.py are not needed
In general, I found it useful to return original datum ids, in case the user wants to load some extra information (regardless of COCO). Part of the problem is the Tuple return type which means any changes are not BC.
Maybe return_dict constructor arg could be introduced which would be more extensible, just like DETR did
cc @pmeier