The Dataset code is currently suffering some inconsistencies in the way things are done leading to
- unneeded review effort
- complex validation
- augmenting the burden of message & documentation translation tasks
Those refactoring direction should include
- all named "x" and "y", and remove "image", "target", ...
- conform to and add the expected class
image_with_bounding_box, image_with_segmentation_mask when relevant