search

Asparagus15 / GaussianShader

code for GaussianShader: 3D Gaussian Splatting with Shading Functions for Reflective Surfaces
Other
288 stars 23 forks source link

How to reproduce the result quality in Fig 2? #7

Closed f1shel closed 9 months ago

f1shel commented 9 months ago

Hi, thanks for this great work! I am reproducing the experiment in Figure 2 with a metallic sphere. I find that the quality of outcome seems have some problem. I am wondering whether I do the things in a right way. Following is the result that I get image Below I describe how I conduct the experiment:

  1. I made a homogenous metallic ball in blender, and rendered it into 120 multi-view images with BlenderNeRF; views are randomly distributed on a sphere.
  2. Gaussian model was optimized on these images with the following command/parameters: 
    python train.py -s dataset/metalball_entiresphere/ --eval -m output/metalball_entiresphere -w --brdf_dim 0 --sh_degree -1 --lambda_predicted_normal 2e-1 --brdf_env 64

    iterations are set to 30000; any other parameters are set to default.

Could you please give some help on the parameters used in this experiment? Thanks in advance!

f1shel commented 9 months ago

Never mind, I get a result very close to the GT sphere just now. The problem disappears with the latest update. It would be better if the normal based on depth derivation can reach the quality of learned normal. Looking forward to the follow-up works.