Towards Real-time Photorealistic 3D Holography with Deep Neural Networks

Project Page | Paper | Data

Liang Shi, Beichen Li, Changil Kim, Petr Kellnhofer, Wojciech Matusik

This repository contains the code to reproduce the results presented in "Towards Real-time Photorealistic 3D Holography with Deep Neural Networks" Nature 2021. Please read the license before using the software.

Getting Started

This code was developed in python 3.7 and Tensorflow 1.15. You can set up a conda environment with the required dependencies using:

conda env create -f environment.yml
conda activate tensorholo

After downloading the hologram dataset, place all subfolders (/*_384, /*_192) into /data directory. The dataset contains raw images and a tfrecord generated for each subfolder. The code by default loads the tfrecord for training, testing and validation.

The current codebase doesn't contain the training code. We will soon make it available in the second phase of release. The current codebase does contain a pretrained CNN for 8um pitch SLMs and code snippet to evaluate the CNN performance on the validation set.

High-level structure

The code is organized as follows:

main.py defines/trains/validates/evaluates the CNN.
optics.py contains optics-related helper functions and various implementations of double phase encoding.
util.py contains several utility functions for network training.
tfrecord.py contains code to generate and parse tfrecord.

Reproducing the experiments

Validate the pretrained model on the validation set

python main.py --validate-mode

Evaluate the pretrained model on arbitrary RGB-D inputs

python main.py --eval-mode

with following options

parser.add_argument('--eval-res-h', default=1080, type=int, help='Input image height in evaluation mode')
parser.add_argument('--eval-res-w', default=1920, type=int, help='Input image width in evaluation mode')
parser.add_argument('--eval-rgb-path', default=os.path.join(cur_dir, "data", "example_input", "couch_rgb.png"), help='Input rgb image path in evaluation mode')
parser.add_argument('--eval-depth-path', default=os.path.join(cur_dir, "data", "example_input", "couch_depth.png"), help='Input depth image path in evaluation mode')
parser.add_argument('--eval-output-path', default=os.path.join(cur_dir, "data", "example_input"), help='Output directory for results')
parser.add_argument('--eval-depth-shift', default=0, help='Depth shift (in mm) from the predicted midpoint hologram to the target hologram plane')
parser.add_argument('--gaussian-sigma', default=0.7, help='Sigma of Gaussian kernel used by AA-DPM')
parser.add_argument('--gaussian-width', default=3, type=int, help='Width of Gaussian kernel used by AA-DPM')
parser.add_argument('--phs-max', default=3.0, help='Maximum phase modulation of SLM in unit of pi')
parser.add_argument('--use-maimone-dpm', action='store_true', help='Use DPM of Maimone et al. 2017')

Citation

If you find our work useful in your research, please cite:

@article{Shi2021:TensorHolography,
author = {L. Shi, B. Li, C. Kim, P. Kellnhofer, W. Matusik},
title = {Towards Real-time Photorealistic 3D Holography with Deep Neural Networks},
journal = {Nature},
issue = {592},
year = {2021},
}

License

Our dataset and code, with exception of the files in "data/example_image", are licensed under a custom license provided by the MIT Technology Licensing Office. By downloading the software, you agree to the terms of this License.

rsn870/tensor_holography