DeepVoxels is an object-specific, persistent 3D feature embedding. It is found by globally optimizing over all available 2D observations of an object in a deeplearning framework. At test time, the training set can be discarded, and DeepVoxels can be used to render novel views of the same object.
This code was developed in python 3.7 and pytorch 1.0. I recommend to use anaconda for dependency management. You can create an environment with name "deepvoxels" with all dependencies like so:
conda env create -f src/environment.yml
The code is organized as follows:
- dataio.py loads training and testing data.
- data_util.py and util.py contain utility functions.
- run_deepvoxels.py contains the training and testing code as well as setting up the dataset, dataloading, command line arguments etc.
- deep_voxels.py contains the core DeepVoxels model.
- custom_layers.py contains implementations of the integration and occlusion submodules.
- projection.py contains utility functions for 3D and projective geometry.
The datasets have been rendered from a set of high-quality 3D scans of a variety of objects. The datasets are available for download here. Each object has its own directory, which is the directory that the "data_root" command-line argument of the run_deepvoxels.py script is pointed to.
- See
python run_deepvoxels.py --help
for all train options. Example train call:
python run_deepvoxels.py --train_test train \
--data_root [path to directory with dataset] \
--logging_root [path to directory where tensorboard summaries and checkpoints should be written to]
To monitor progress, the training code writes tensorboard summaries every 100 steps into a "runs" subdirectory in the logging_root.
Example test call:
python run_deepvoxels.py --train_test test \
--data_root [path to directory with dataset] ]
--logging_root [path to directoy where test output should be written to] \
--checkpoint [path to checkpoint]
If you find our work useful in your research, please consider citing:
@inproceedings{sitzmann2019deepvoxels,
author = {Sitzmann, Vincent
and Thies, Justus
and Heide, Felix
and Nie{\ss}ner, Matthias
and Wetzstein, Gordon
and Zollh{\"o}fer, Michael},
title = {DeepVoxels: Learning Persistent 3D Feature Embeddings},
booktitle = {Proc. CVPR},
year={2019}
}
The code in the subdirectory "pytorch_prototyping" comes from a little library of custom pytorch modules that I use throughout my research projects. You can find it here.
Some of the code in this project is based on code from these two very cool papers: Learning a Multi-View Stereo Machine 3DMV Check them out!
We have more cool work on novel view synthesis, neural scene representations and neural rendering in the pipeline - stay tuned!
If you have any questions, please email Vincent Sitzmann at [email protected].