TT-NF: Tensor Train Neural Fields

This repository implements experiments from our paper titled "TT-NF: Tensor Train Neural Fields" by Anton Obukhov, Mikhail Usvyatsov, Christos Sakaridis, Konrad Schindler, and Luc Van Gool. [Project Website]

Abstract

Learning neural fields has been an active topic in deep learning research, focusing, among other issues, on finding more compact and easy-to-fit representations. In this paper, we introduce a novel low-rank representation termed Tensor Train Neural Fields (TT-NF) for learning neural fields on dense regular grids and efficient methods for sampling from them. Our representation is a TT parameterization of the neural field, trained with backpropagation to minimize a non-convex objective. We analyze the effect of low-rank compression on the downstream task quality metrics in two settings. First, we demonstrate the efficiency of our method in a sandbox task of tensor denoising, which admits comparison with SVD-based schemes designed to minimize reconstruction error. Furthermore, we apply the proposed approach to Neural Radiance Fields, where the low-rank structure of the field corresponding to the best quality can be discovered only through learning.

Setting Up Environment

Experiments can be reproduced on a single 16Gb GPU.

Clone the repository, then create a new virtual environment, and install python dependencies into it:

python3 -m venv venv_ttnf
source venv_ttnf/bin/activate
pip3 install --upgrade pip
pip3 install -r requirements.txt

Experiments

Tensor Denoising

Run the following command:

python -m src.exp_ttnf.tensor_denoising

Neural Rendering

Choose a preconfigured experiment from src/exp_qttnf/configs directory, and run the following command:

CUDA_VISIBLE_DEVICES=0 python -m src.exp_qttnf.run_qttnf_nerf \
    --config src/exp_qttnf/configs/<config.txt> \
    --log_root <path_to_artifacts_dir> \
    --dataset_root <path_to_dataset_dir>

Logging

The code performs logging to the console, tensorboard file in the experiment log directory, and also Weights and Biases. Upon the first run, please enter your account credentials, which can be obtained by registering a free account with the service.

Citation

Please cite our work if you found it useful:

@misc{obukhov2022ttnf,
  author = {Obukhov, Anton and Usvyatsov, Mikhail and Sakaridis, Christos and Schindler, Konrad and Van Gool, Luc},
  title = {TT-NF: Tensor Train Neural Fields},
  year = {2022},  
  doi = {10.48550/ARXIV.2209.15529},
  url = {https://arxiv.org/abs/2209.15529},
  publisher = {arXiv},
  copyright = {Creative Commons Attribution Share Alike 4.0 International}
}

License

This software is released under a CC-BY-NC 4.0 license, which allows personal and research use only. For a commercial license, please contact the authors. You can view a license summary here.

Portions of source code are taken from external sources under different licenses, including the following:

• NeRF (MIT)
• TTOI (rewritten from MatLab)
• Spherical Harmonics (BSD-2)

Acknowledgements

This work was supported by Toyota Motor Europe and was carried out at the TRACE Lab at ETH Zurich (Toyota Research on Automated Cars in Europe - Zurich).