search

liuyuan-pal / NeRO

[SIGGRAPH2023] NeRO: Neural Geometry and BRDF Reconstruction of Reflective Objects from Multiview Images
MIT License
537 stars 35 forks source link

About the rescaling of roughness in Stage 2. #42

Open K-nowing opened 8 months ago

K-nowing commented 8 months ago

Thank you for your work!

From the code of stage 2, why do you set the min of roughness 0.04? https://github.com/liuyuan-pal/NeRO/blob/7f3673532113cf40433d4d686c9ac7e8c0f9813f/network/field.py#L919

Is there a physical reason for this, or is it just a hyperparameter?

Thanks!

liuyuan-pal commented 8 months ago

Hi, this is only a hyperparameter. When I use 0, there may be some numerical issues when computing the BRDF and there are no perfectly smooth objects. Thus, I chose this 0.04 threshold. Thank you!

K-nowing commented 8 months ago

Thank you!