PIL.UnidentifiedImageError when downloading dataset with missing image

[stevensagaert]

Due to the fact that currently there is no transactional protection between metadata en image data it's possible to get a corrupted dataset where the metadata has an image path and marks a sample as annotated & reviewed but there is no actual image. When you try to download such a dataset you get a PIL.UnidentifiedImageError.

The following modified dataset.py has the necessary minimal changes to be able to download such a dataset but of course this solution isn't fully satisfactory because the client still has to deal with corrupt data (in this case missing images). The real good solution would be to prevent such corruption in the first place. Note that even in this case some changes to dataset.p error handling are still useful because some setting of dummy metadata in case of an error just didn't work (in wrong branch).
Note that the changes also include the change for issue #68
dataset.zip