2D-to-3D lifting issues with objects outside PartNetE #10
Description
Hello,
I’m encountering a problem with the 2D-to-3D lifting part of the system, with both semantic and instance segmentation. I’ve tried on both PartNet chairs (w/o texture) converted to .ply and then sampled to make a pointcloud, and a different chair model outside PartNetE (w/ texture, since issue #4 said texturing might be an important factor); the GLIP and SAM predictions look pretty reasonable (attached some examples for the textured chair below), but the resulting segmentation .plys are often completely black because there are no coherent regions that have been assigned a label.
Do you have ideas on what might be going wrong? If this is a known limitation or bug, are there any recommended workarounds or debug steps?
Thanks!
GLIP/SAM result examples
All resulting segmentation .plys look like this:
