About Monocular Depth or Normal Loss

[buddyhs]

Thank you for your excellent work. The training results are very good. I have another question. Since PGSR can already render depth maps and normal vectors, can the current code add single-purpose depth or normal losses to constrain more accurate model generation?

[ywcmaike]

It depends on the generalization ability of the pretrained depth or normal models. From my perspective, in the texture-less areas or some ambiguous areas, using the prior of the pretrained normal or depth model may help improve the quality of the surface reconstruction.