duplicated annotation ids in the validation dataset

[Gary Fan]

I found the annotation ids for bbox annotations in validation/panoptic/panoptic_2018.json seem to have duplicates. This will lead to issues when converting to coco instance format. Wonder if they are intentional and if there are specific reasons to keep it that way?

len([item['id'] for item in segment_infos]), len(set([item['id'] for item in segment_infos])) len([item['id'] for item in segment_infos]), len(set([item['id'] for item in segment_infos])) (136031, 102954)

Much appreciated!

[Peter Kontschieder]

@garyfanhku The ID's are unique, but this is only required on a per-image basis, and not over the entire {training,validation,test} sets.
Please take a look at the label description here, and you can see that an annotation exists on a per-image basis, holding its own segments_info. For this reason, we do not require IDs to be unique globally. HTH!