Part 2 of trying #SegmentAnythingModel (#SAM) on capsule endoscopic data from #KvasirCapsule: From bounding boxes to uncertain segmentation masks.

What you see in the pictures are two sample images from #KvasirCapsule, classified by the clinical experts as an erosion and annotated with a bounding box around the lesion. From within the bbox, we took N=500 sample points, with which we prompted SAM.

I just tried #SegmentAnythingModel #SAM on capsule endoscopic images from #KvasirCapsule.

Mask generation on full image over-segments a lot, which was expected.

See alt text for more details: