TensoRF
TensoRF copied to clipboard
Arguments to the model
Thank you for the code. Could you please add some comments what the arguments to the model mean? in TensorBase.
Hi @SrinjaySarkar - I read the paper and have been digging around this repo for a while. Are there any args in particular which you're not clear on? I could try to provide detail (with references of course).
Thank you Zain!
Hi @SrinjaySarkar - I read the paper and have been digging around this repo for a while. Are there any args in particular which you're not clear on? I could try to provide detail (with references of course).
Hello.If you could answer my doubts, I would greatly appreciate it.
1、I used my own phone to capture the dataset, so I should follow 'own_data 'type(ues 'colmap2nerf.py') or generate according to' llff 'type(use 'imgs2poses.py'). Will these two formats affect subsequent results? Because I see that the configuration file parameters of the two are different, why and how should I choose
2、The author said that it is not suitable to generate content with a background. How does the algorithm distinguish the background?
3、Why do you think there are some issues with using my own dataset for training and generating depth maps
4、How to understand parameters“update_AlphaMask_list”、“n_lamb_sigma ”、“n_lamb_sh ”
Hi @jly0810 - to answer your questions 1, 3, and 4 above - to be frank, my memory of this repo has gotten a little dusty. So my honest recommendation might be to switch to using another implementation of TensoRF (such as that found in NeRFStudio), which has been optimized to be alot more understandable and user-friendly.
For your question 2 - I'm actually not quite sure! My (superficial) understanding is that TensoRF technically can capture scenes that include a background (e.g., we can use the Tanks and Temples dataset); but if your goal is to render novel viewpoints of the background, then that's when you'd start getting lots of floater artifacts (and I'd be curious to know what @apchenstu thinks is the reason why, on a deeper mathematical level). At that point, you might want to switch to another method. You could probably ask around in the NeRFStudio Discord for recommendations on which to go with - my personal guess is to perhaps go with Mip-NeRF. Hope that helps!