TokenCut false positives when there is no object in the image

[surajbijjahalli]

Im trying to apply TokenCut to a dataset in which there is a combination of images without salient objects (flat, relatively textureless images), and images in which there are clear objects (objects we are interested in). Is there a way of reducing false positives or some metric to indicate that the foreground segment is a weak one? or will TokenCut always return a segmentation even when there is no clear object in the scene ?

[XiSHEN0220]

Hi, nice question!

Currently, we don't have any strategies to filter the false positives samples.
I do think it would be a very good research direction.

Best,

Xi

[surajbijjahalli]

Thank you Xi. Will post here when I find a good strategy !