Hollo, I've read your paper and the source code, and have a question about the $f^{geo}$ branch of the dual color network. The code lib/voxurf_womask_fine.py line 949
rgb_logit = self.rgbnet(rgb_feat)
shows that the rgb_feat, i.e. the neighbouring SDF values and their differences require gradients, so that SDF values are optimized through the color predicting branch. Is that reasonable?
And when I add a detach() operation to rgb_feat, the CD in DTU24 increased from 0.7+ to 1.0+. Following is the comparisons.
A straightforward guess is that the $f^{geo}$ branch indicates that similar geometry patterns generate similar color (like roof in DTU24), therefore optimizing the SDF patterns of roof consistently, leading to smooth roof surfaces.
Hollo, I've read your paper and the source code, and have a question about the $f^{geo}$ branch of the dual color network. The code
lib/voxurf_womask_fine.py
line 949shows that the rgb_feat, i.e. the neighbouring SDF values and their differences require gradients, so that SDF values are optimized through the color predicting branch. Is that reasonable?
And when I add a
detach()
operation torgb_feat
, the CD in DTU24 increased from 0.7+ to 1.0+. Following is the comparisons.A straightforward guess is that the $f^{geo}$ branch indicates that similar geometry patterns generate similar color (like roof in DTU24), therefore optimizing the SDF patterns of roof consistently, leading to smooth roof surfaces.
===============================================================
Maybe should keep the gradients of
k0
. So I tried to only detachall_feat
,all_grad
andcenter_sdf
(line 931, 932 and 947). But the CD is still 1.0+.