Binary Radiance Fields (NeurIPS 2023)

Paper | Project Page

This repository contains the code release for the paper:

Binary Radiance Fields
Seungjoo Shin¹, and Jaesik Park^2
1 POSTECH, ² Seoul National University
Conference on Neural Information Processing Systems (NeurIPS), New Orleans, 2023

Setup

We have tested on PyTorch==1.13.0 with CUDA==11.7.

Clone the repository:

git clone https://github.com/SAITPublic/BiRF.git
cd BiRF

Create a environment:

conda create --name birf -y python=3.8
conda activate birf

Install packages:

# install PyTorch==1.13.0 with CUDA==11.7
conda install pytorch==1.13.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.7 -c pytorch
# install custom tiny-cuda-nn
pip install git+https://github.com/seungjooshin/tiny-cuda-nn/@bit#subdirectory=bindings/torch
# install requirements
pip install -r requirements.txt

Prepare datasets:

We support three datasets for evaluation.

• Synthetic-NeRF
• Synthetic-NSVF
• Tanks and Temples

By default, we put the dataset under ./data/ as:

├── data
├── nerf_synthetic
├── chair
├── drums
├── ....
├── Synthetic_NSVF
├── Bike
├── Liferstyle
├── ....
├── TanksAndTemples
├── Barn
├── Caterpillar
├── ....

Running

Training:

# python train.py ./config/{size}.gin --scene {scene} --n_features {n_features}
python train.py ./config/base.gin --scene chair --n_features 2

• size: the size of hash table
• scene: the scene to reconstruct
• n_features: the number of features
• The result is saved as {log_dir}/results_{seed}.json.

Testing:

# python test.py ./config/{size}.gin --scene {scene} --n_features {n_features} --log_dir {path_to_log_dir}
python test.py ./config/base.gin --scene chair --n_features 2 --log_dir ./logs/chair_f2_2023

• size: the size of hash table
• scene: the scene to reconstruct
• n_features: the number of features
• log_dir: the path to log directory
• The result is saved as {log_dir}/results.json.

By default, we save the log under ./logs/.

├── logs
├── chair_b_2
├── imgs
├── 0000.png
├── 0001.png
├── ....
├── config.gin
├── encoding.npz
├── network.ckpt
├── occgrid.npz
├── results.json
├── chair_s_2
├── ....

Citation

If you find our code or paper useful, please consider citing our paper:

@inproceedings{
shin2023binary,
title={Binary Radiance Fields},
author={Seungjoo Shin and Jaesik Park},
booktitle={Advances in Neural Information Processing Systems},
year={2023}
}