This code.zip archive provides the raw code for the paper "Efficient Unrolled Networks for Large-Scale 3D Inverse Problems", Romain Vo and Julián Tachella.

The code is divided into three parts:
- a specific deepinv fork, which adapts some of the library's functionalities for the paper
- unrolled3d-cvpr/scripts, which provides the raw code and scripts for running the experiments in the paper (it only works with the provided deepinv fork)
- inr-3d-ct, used to compute the instant-NGP reconstructions for CBCT (the code is identical to https://github.com/romainvo/nef-red)

The CBCT dataset can be download here: https://zenodo.org/records/2686726. This dataset is composed of 42 .zip files. Each .zip contains .tiff files for the projections as well as already computed reconstructions, it also contains the trajectory of the source, detector and object during the acquisition. Once all the .zip files have been downloaded and placed in <data> folder, it can be processed using the prepare_ct_data.py by replacing the variable input_dir with your <data> location. This script will assemble the .tiff files into .raw files easier to manipulate. In the <data> dir you can then run the create_ct_data_splits.py script, it will create the train/val/test splits as well as a dataset_50p.csv files to store useful informations on the geometry, and sample easily from the dataset during the experiment. The data is split into 30 train, 4 val, 8 test.

The MC-MRI dataset can be downloaded here: https://sites.google.com/view/calgary-campinas-dataset/download. The sampling mask are available at this repository: https://github.com/rmsouza01/MC-MRI-Rec/tree/master/Data/poisson_sampling.

UPDATE: For weights and an easy-to-use version of our work, check out our updated repo: https://github.com/romainvo/efficient-unrolling !
It provides press-button functionality to integrate any deepinv physics into our domain partitioning framework.