Leolty / NeRF-PyTorch

A pytorch version of NeRF

NeRF: Neural Radiance Fields for Novel View Synthesis

A NeRF re-implementation using PyTorch - Machine Learning for 3D Data, CSE291 22Fall UCSD

---

written by Tianyang Liu

Quick Start

First, create a conda enviroment named nerf and install corresponding dependencies with the following bash code.

conda env create -f environment.yml
conda activate nerf

Then, open workflow.ipynb, just follow the steps.

NeRF Pipeline

1. Train a single NeRF with data argumentation
2. Drop data argumentation and continue to train the single NeRF.
3. Hard copy the single NeRF to two (one for coarse, one for fine), and use hierarchical samping to fine the outpit.
4. Finetune NeRF system.

Rendering Effect

Reference

The implementation is based on some Github repos (nerf, NeRF, nerf_pl and nerf-pytorch), the example of nerf on keras, some Google Colabs (NeRF.ipynb and NeRF_From_Nothing), and the article on Medium.

Of course, the present implementation also draws heavily on NeRF original paper.