Open slothfulxtx opened 10 months ago
Sorry for the delay and thanks for the attention. Please check the latest update. I have uploaded an example data with training and testing instructions. The hyperparameter default values are also updated.
Hi, thanks for your great work and the clever idea of the short axis normal.
However, when I tried to replicate the results of the paper, I encountered some inconsistencies. I cloned the latest code, and my command for running it was shown below.
python train.py -s ../data/refnerf/ball --eval -m output/ball_blender -w --brdf_dim 0 --sh_degree -1 --lambda_predicted_normal 2e-1 --brdf_env 512
This resulted in the following outcomes.
Actually, for me, I was hoping to obtain a relatively perfect normal. It would be great if the author could point out the core idea behind the training of the normals, such as the learning rate, supervised loss, etc.
Hi, authors, Thanks for your interesting work! When I tried to reproduce the performance on Shiny Blender dataset and run your code, I found that some of your default arguments are not consistent with your paper. For example, the weighted hyperparameters of each loss terms. After I fixed these issues and made the code runnable, I still cannot reproduce the performance you report in the paper. Could you tell me how to config the arguments for shiny blender? For examples,
brdf_dim = ?