This is the official implementation of VaxNeRF (Voxel-Accelearated NeRF).
VaxNeRF provides very fast training and slightly higher scores compared to original (Jax)NeRF!!

• December 26, 2021
  • We Vax'ed MipNeRF!
  • We achieved roughly 28 times faster training in terms of producing the final accuracy of the original NeRF.
  • See -> https://github.com/naruya/mipnerf
• December 2, 2021
  • We Vax'ed PlenOctrees! You can train NeRF-SH about 5x faster.
  • https://github.com/naruya/plenoctree (out of maintenance)

Visual Hull (1sec)
NeRF (10min)
VaxNeRF (10min)
Vax-MipNeRF (10min)

(The results of Vax-MipNeRF are also included in this figure.)

Please see the README of JaxNeRF.

The jax and jaxlib versions that we have tested are as follows.

jax                     0.2.24
jaxlib                  0.1.69+cuda111

jax                     0.2.17
jaxlib                  0.1.65+cuda110

# make a bounding volume voxel using Visual Hull
python visualhull.py \
    --config configs/demo \
    --data_dir data/nerf_synthetic/lego \
    --voxel_dir data/voxel_dil7/lego \
    --dilation 7 \
    --thresh 1. \
    --alpha_bkgd

# train VaxNeRF
python train.py \
    --config configs/demo \
    --data_dir data/nerf_synthetic/lego \
    --voxel_dir data/voxel_dil7/lego \
    --train_dir logs/lego_vax_c800 \
    --num_coarse_samples 800 \
    --render_every 2500

python eval.py \
    --config configs/demo \
    --data_dir data/nerf_synthetic/lego \
    --voxel_dir data/voxel_dil7/lego \
    --train_dir logs/lego_vax_c800 \
    --num_coarse_samples 800

Original NeRF

python train.py \
    --config configs/demo \
    --data_dir data/nerf_synthetic/lego \
    --train_dir logs/lego_c64f128 \
    --num_coarse_samples 64 \
    --num_fine_samples 128 \
    --render_every 2500

VaxNeRF with hierarchical sampling

# small `num_xx_samples` needs more dilated voxel (see our paper)
python visualhull.py \
    --config configs/demo \
    --data_dir data/nerf_synthetic/lego \
    --voxel_dir data/voxel_dil47/lego \
    --dilation 47 \
    --thresh 1. \
    --alpha_bkgd

# train VaxNeRF
python train.py \
    --config configs/demo \
    --data_dir data/nerf_synthetic/lego \
    --voxel_dir data/voxel_dil47/lego \
    --train_dir logs/lego_vax_c64f128 \
    --num_coarse_samples 64 \
    --num_fine_samples 128 \
    --render_every 2500

Visual Hull

• Use --dilation 11 / --dilation 51 for NSVF-Synthetic dataset for training VaxNeRF without / with hierarchical sampling.
• The following options were used
• Since the Lifestyle, Spaceship, Steamtrain scenes (included in the NSVF dataset) do not have alpha channel, please use following options and remove --alpha_bkgd option.
  • Lifestyle: --thresh 0.95, Spaceship: --thresh 0.9, Steamtrain: --thresh 0.95

NeRFs

• We used --small_lr_at_first option for original NeRF training on the Robot and Spaceship scenes to avoid local minimum.

• You can see the main difference between (Jax)NeRF (jaxnerf branch) and VaxNeRF (vaxnerf branch) here
• The main branch (derived from the vaxnerf branch) contains the following features.
  • Support for original NeRF
  • Support for VaxNeRF with hierarchical sampling
  • Support for the NSVF-Synthetic dataset
  • Visualization of number of sampling points evaluated by MLP (VaxNeRF)
  • Automatic choice of the number of sampling points to be evaluated (VaxNeRF)

Please use the following bibtex for citations:

@article{kondo2021vaxnerf,
  title={VaxNeRF: Revisiting the Classic for Voxel-Accelerated Neural Radiance Field},
  author={Kondo, Naruya and Ikeda, Yuya and Tagliasacchi, Andrea and Matsuo, Yutaka and Ochiai, Yoichi and Gu, Shixiang Shane},
  journal={arXiv preprint arXiv:2111.13112},
  year={2021}
}

and also cite the original NeRF paper and JaxNeRF implementation:

@inproceedings{mildenhall2020nerf,
  title={NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis},
  author={Ben Mildenhall and Pratul P. Srinivasan and Matthew Tancik and Jonathan T. Barron and Ravi Ramamoorthi and Ren Ng},
  year={2020},
  booktitle={ECCV},
}

@software{jaxnerf2020github,
  author = {Boyang Deng and Jonathan T. Barron and Pratul P. Srinivasan},
  title = {{JaxNeRF}: an efficient {JAX} implementation of {NeRF}},
  url = {https://github.com/google-research/google-research/tree/master/jaxnerf},
  version = {0.0},
  year = {2020},
}

We'd like to express deep thanks to the inventors of NeRF and JaxNeRF.

Have a good VaxNeRF'ed life!