search

bmild / nerf

Code release for NeRF (Neural Radiance Fields)
http://tancik.com/nerf
MIT License
9.84k stars 1.37k forks source link

Why no_batching=True when using blender data type? #108

Open sstzal opened 3 years ago

sstzal commented 3 years ago

Hi,

I'm a little confused about the config setting: Why no_batching=True when using blender data type? Are there any special considerations for doing this?

I notice that when no_batching=True, the code will sample rays near the center of the image, and I'm confused about why doing such operation with blender data rather than llff data type?

Thanks

tancik commented 3 years ago

The terminology is a little confusing. When no_batching is set to true, during each iteration the entire batch of rays are only sampled from a single image. When it is false, we sample the rays from all of the images during each iteration. Setting it to False will improve the results, but it will also consume more memory in the current implementation. Because there are more images in the synthetic scenes, we needed to set this value to True to avoid OOM errors. With smarter data loading and better memory management this can easily be avoided, some of the other NeRF reimplementation fix this issue. In order to match the paper results, we have decided not to make these modifications.