This repository contains the code to reproduce the results presented in "Three-dimensional imaging through scattering media based on confocal diffuse tomography" by David B. Lindell and Gordon Wetzstein.
Code prerequisites: Linux-based operating system, installation of Anaconda or Miniconda. Developed and tested with Python 3.6 on Arch Linux kernel 5.6.8-arch1-1 with an Intel Core i7-9750H CPU.
The code is developed in Pytorch and supports both CPU and GPU execution. Most NVIDIA GPUs should be automatically detected and used.
List of contents
main.py- wrapper program called to generate results
cdt_reconstruction.py- contains processing code to run the reconstructions and reproduce results of main paper.
data/cones.mat- data file for Fig. 3
data/letter_s.mat- data file for Fig. 2
data/letters_ut.mat- data file for Fig. 3
data/letter_t.mat- data file for Fig. 3
data/letter_u*.mat- data files shown in Supplementary Movie 2 for an object at varying distances behind the media
data/mannequin.mat- data file for Fig. 3
data/resolution*.mat- data files for Supplementary Fig. 17
README.txt- this file
requirements.txt- list of requisite Python packages
setup.bash- example script to install a conda environment and required packages and run the code
utils.py- contains helper functions for the processing
To run the demo code, follow the instructions in the setup.bash script. This script assumes that you have Anaconda or Miniconda installed. If not, follow the provided link for instructions on how to do this. The script sets up a new Python 3.6 environment, installs the required Python packages (listed in
requirements.txt), and runs the reconstruction code.
The install time of the setup.bash script is less than 5 minutes on the tested configuration. Runtime of the demo program
main.py is less than one minute on the tested configuration and outputs.
The expected output of the
main.py program is a figure showing maximum intensity projections of the selected 3D measurement volume and reconstruction.
python main.py --help for a list of commandline options for running the demo.
Code to estimate the parameters of the scattering medium is included in the
calibration folder. The data capture and optimization procedure are described in the Methods and Supplementary Information of the paper. We jointly optimize the parameters over captured time-resolved transmittance profiles for varying thicknesses of the scattering material (from 1 to 8 inches). Since the optimization is non-linear, we vary the initialized values of the reduced scattering coefficient and refractive index and select the values that best fit the data. Analyzing the distribution of optimized values is also useful for understanding the variance of the resulting estimate.
sweep_n.py- script to run the optimizations over all initializations
optimize.py- main optimization script
diffusion_model.py- wrapper class that contains the diffusion model and loads and processes captured measurements