GH-NeRF

---

“GH-NeRF”, is a NeRF-like model which uses 3D multivariate Gaussian random variables along with Spherical Harmonics to accelerate the training process by 25% and also, tackle the problem of excessively blurring or alias of the original NeRF implementation, lowering the error rate by 6.4% relative to NeRF on LLFF dataset.

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

Architecture

Installation/Train:

Preliminaries:

cd GH-NeRF
pip 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

Result

trex:

Mini paper

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

Reference

---

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