Limit mask size in Mask2Former results

🤗Transformers
1

Hi,

I use Mask2Former for instance segmentation with direct integration in CVAT using the standard pipeline.

I noticed that Mask2Former for instance segmentation tends to detect :

  • Lots of small objects (a few pixels) that are not interesting.
  • Sometimes, the mask of an object vampirizes a big part of the background.

It’s not only our models; the demo models from Facebook also do that.

Claude suggested adding “mask_size_range”: [0.05, 0.80] in config.json, but I see no effect in CVAT. I did not find this in the doc…

I can’t add a filter, as CVAT automatic integration of Mask2Former has to use the standard pipeline, or it means that I have to make a custom integration (which I would like to avoid for the time being).

Is there a way to prevent Mask2Former from creating very small and very big masks?

For big masks, I’m wondering if I shall switch to a panoptic Mask2Former so that he learns the background better. For small masks, I have no ideas unless there is some parameter for that.

Best regards

1 Like
2

It seems that you need to manually filter out small rectangular areas. However, as the post is old, it is possible that it has been improved since then.

1 Like