search

lojzezust / lars_evaluator

Evaluation scripts for the LaRS dataset
https://lojzezust.github.io/lars-dataset
10 stars 2 forks source link

How can we convert the panoptic annotations to coco instance.json format? #4

Open yaoshanliang opened 1 month ago

yaoshanliang commented 1 month ago

The original Mask2Former repo requires instance.json labels. However, I faced challenges when I tried to convert panoptic.json to instance.json. Could you please provide the instance labels, or do you know how to do the convertion? Thanks in advance.

lojzezust commented 1 month ago

It should be possible to convert labels in the COCO panoptic format (JSON + mask PNGs) to the COCO instance format. Details about the formats can be found here. The difficult part might be RLE encoding the masks (only the part inside the bbox).

However, I believe Mask2Former should support training from panoptic annotations only. Note that sometimes instances.json are used for evaluation only, and not needed for training.

You can also take a look at the mmdetection implementation of Mask2Former. For that one I know it's possible to train without instance JSON files. This fork even contains some basic configs and conversion scripts for LaRS, you might find useful.