About NeRF Dataset

[2j1ejyu]

Thanks for the great repo!

If I understood the code correctly, the NeRF Dataset in your code (nerf/provider.py) is getting each batch from the same image rays, randomly sampling, which is different from the original NeRF Dataset.
I want to ask if there's a reason you coded this way and also want to ask if there would be any performance degradation

1. since training with rays in the same image could bother training the scene generally(this is just my opinion) and
2. since the rays are sampled randomly, some rays could be less trained

[ashawkey]

@2j1ejyu Hi, these two behaviours in fact follow instant-ngp. In some early experiments I don't find these will cause significant performance degradation. You can modify them and try!

[2j1ejyu]

thank you so much for your quick reply! Have nice day :)