The depth estimation capability of the NESTED-GIANT-LARGE model degrades significantly on no-background samples (e.x. pure black or white background), which affects camera pose estimation and 3D object generation.
In addition, I found that the depth estimation of DA3 has degradation on the surfaces of transparent and reflective objects, as well as on some fine structures.
I wonder if anyone has encountered the same problem as I have.