point SAM performance on interior boundaries

[jeb2112]

I have run the demo app.py on the rhino point cloud, and I am seeing results that are essentially equivalent to the rhino demo video shown on the github.io page (which appears to be taken with a more refined version of the app.py). However, the demo videos seem to focus on segmentations that have a combination of interior and exterior boundaries. So, for example in the rhino's legs or tail or horns, most of the intended segmentation boundaries are exterior air boundaries, and a smaller interior boundary is only present at the chest, or head or rump. In my application, the point cloud is not hollow, and further I have predominantly interior boundaries. I created a simple high-contrast 3d object out of squares and rectangles with mostly interior boundaries to test this, and I am seeing that the segmentations are not so good. Is there any indication of how well point SAM could, or should, work on dense point clouds with interior boundaries?