Closed jpsml closed 3 months ago
Our method does not model different roughness level and interreflections. Our method tends to first fit reflections with higher roughness, so the env. map looks a little bit blurred. Additionally, interreflections can affect normal reconstruction, resulting in poorer performance in the material scene.
I executed the following experiment:
python -u train.py -s data/nerf_synthetic/materials/ --eval --iterations 61000 --white_background --densification_interval_when_prop 100
However, the obtained normal maps are a little bit noisy in the spheres with high specularity:
I noticed that the obtained env maps are a little bit blurred:
Is this the expected result?