mipnerf-pytorch

This repository contains the code release for Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. This implementation is written in pytorch, and is an unofficial pytorch implement of of Google's mipnerf implementation.

Result

	Multi Scale Train And Multi Scale Test
	PNSR						SSIM
	Full Res	1/2 Res	1/4 Res	1/8 Res	Aveage (PyTorch)	Aveage (Jax)	Full Res	1/2 Res	1/4 Res	1/8 Res	Average (PyTorch)	Average (Jax)
lego	34.74	35.93	36.29	35.62	35.65	35.74	0.9719	0.9841	0.9890	0.9894	0.9836	0.9843

The above results are trained on the lego dataset with 750k (Training not completed yet and 1000k steps in total) steps for multi-scale datasets, and the pre-trained model can be found here with password psrd.

Installation

We recommend using Anaconda to set up the environment. Run the following commands:

# Clone the repo
git clone https://github.com/AlphaPlusTT/mipnerf-pytorch.git; cd mipnerf-pytorch
# Create a conda environment, note you can use python 3.6-3.8 as one of the dependencies.
conda create --name mipnerf python=3.8.11; conda activate mipnerf
# Prepare pip
conda install pip; pip install --upgrade pip
# Install requirements
We used torch==1.9.1, torchvision==0.10.1, CUDA==10.2, hydra-core==1.1.1, visdom==0.1.8.9, matplotlib==3.5.1, einops==0.4.1

Data

Then, you'll need to download the datasets from the NeRF official Google Drive. Please download and unzip nerf_synthetic.zip and nerf_llff_data.zip.

Generate multiscale dataset

You can generate the multiscale dataset used in the paper by running the following command,

python scripts/convert_blender_data.py --blenderdir /nerf_synthetic --outdir /multiscale

Running

Just modify the config file and

python trian.py

OOM errors

You may need to reduce the batch size to avoid out of memory errors. For example the model can be run on a NVIDIA 3080 (10Gb) using the following yaml.

train:
  batch_size: 1024

Citation

Kudos to the authors for their amazing results:

@misc{barron2021mipnerf,
      title={Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields},
      author={Jonathan T. Barron and Ben Mildenhall and Matthew Tancik and Peter Hedman and Ricardo Martin-Brualla and Pratul P. Srinivasan},
      year={2021},
      eprint={2103.13415},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgements

Thanks to Boyang Deng for JaxNeRF. Thanks to jonbarron for mipnerf. Thanks to facebookresearch for nerf.