19reborn/NeuS2

Color gets worse in training

xmeng525 opened this issue · 3 comments

Dear authors,
Thank you for the great work. I have a question about the color of exported mesh.
I tried to reconstruct a sphere with 26 viewpoints. Here are 4 example views of the input.
000023000018000008000000

In the training process, everything loos normal (the RGB loss and the eikonal loss decreases; the geometry get better; and the rendered images look good). However, I get more and more noise in the vertex color of the exported mesh as shown below.
Epoch = 500
ep50003
Epoch = 1500 (noise starts to appear)
ep150004
Epoch = 3000 (the colors becomes dark and gloomy)
ep300005

Have you observed similar phenomenon in your experiments? Do you have any clue about the reasons?

sphere.zip
I have attached the dataset in case you would like to take a look. The configuration file I used is configs/nerf/base.json.
Thanks!

Hi @xmeng525 , thanks for using our codes! I guess the main reason is that the image resolution is too low (64x64).

Thank you for answering. Looking forward to seeing more great work from U!