Repository files navigation

“GH-NeRF”, is a NeRF-like model which uses 3D multivariate Gaussian random variables along with Spherical Harmonics to accelerate the training process by25% and also, tackle the problem of excessively blurring or alias of the original NeRF implementation, lowering the error rate by6.4% relative to NeRF on LLFF dataset.

TL;DR: Mip-NeRF + NeRF-SH (promoted by Spherical Harmonics) implementation in Pytorch

Preliminaries:

cd GH-NeRFpip install -r requirements.txt

To download dataset:

• bash scripts/download_llff.sh to download LLFF

To train the model:

python run_nerf.py --config configs/trex.txt

Project Layout:

├─configs├─data│  ├─nerf_llff_data│     ├─fern│     │  ├─images│     │  ├─images_4│     │  ├─images_8│     │  ├─mpis4│     │  └─sparse│     │      └─0│     └─trex│         ├─images│         ├─images_4│         ├─images_8│         ├─outputs│         └─sparse│             └─0├─logs│  ├─fern_test│  │  └─train│  └─trex_test│      ├─testset_200000|      ├─...│      └─train├─scripts├─static

trex:

see./static/paper.pdf in the project.

1. NeRF
2. Mip-NeRF
3. MINE: Continuous-Depth MPI with Neural Radiance Fields
4. PlenOctrees
5. Plenoxels
6. nerf-pytorch
7. mipnerf-pytorch