Questions about ncc_scale and DTU results

[Zzzzzzy123]

Hello, thanks for your great work! However, I observed that PGSR sets ncc_scale=0.5 when evaluating CD on the DTU dataset. However, this configuration employs full-resolution grayscale images for NCC loss calculation, unlike most methods (2dgs) that use half-resolution images as ground truth. I think this will lead to potential unfair advantages, as the higher-resolution input may leak additional detail to PGSR during optimization.

To investigate, I retested PGSR on some DTU scenes with ncc_scale=1. The results show that while some scenes exhibit CD scores consistent with the paper, others demonstrate significant degradation in metrics.

	scan37	scan97	scan122
paper	0.52	0.69	0.36
ncc_scale=1	0.67	0.69	0.40

is this a mistake or misunderstanding? Could you please clarify？

PGSR/scene/cameras.py

Line 39 in de24f1a

ncc_resolution = (int(resolution[0]/ncc_scale), int(resolution[1]/ncc_scale))

PGSR/scene/cameras.py

Line 46 in de24f1a

ncc_resolution = (int(resolution[0]/ncc_scale), int(resolution[1]/ncc_scale))

[turandai]

Hi, I was also confused by the ncc_scale, have you tested on other scenes? I wonder how important it is.