distance_scale 25

[w-hc]

Hi thanks a lot for releasing the code for the nice work

distance_scale=25 is used at train and test time for rendering, but not for alpha mask update or mesh extraction. I can't seem to find discussion regarding this hyperparameter in the main text, and I wonder whether you could provide a little more explanations.

It appears that using distance_scale for mesh extraction results in noisier geometry with more floaters, while not applying distance_scale during training would lead to divergence. Thanks again.

[apchenstu]

Hi, thanks! It's equivalent to mod with distance_scale and adjusts the thresholds in building the binary volume or marching cude

[gaoxiangjun]

Hi, could you explain that why not applying distance_scale during training would lead to divergence ?

[Xiaoming-Zhao]

Hi @apchenstu, I tried to disable distance_scale and density_shift during training. Namely, I set distance_scale = 1 and density_shift = 0. After 30k-iter training, this gives me test PSNR of 33.93 on the Lego scene, which is quite inferior to the 36.46 reported in the paper.

May I know whether you mind sharing some intuition why we need such a scale and shift during training to boost the performance? Thanks a lot.

@w-hc , I think it somehow requires to also change density_shift if you disable distance_scale. I also encounter issues if I do not change density_shift, i.e., all alpha values are nearly zeros for all voxels.