Source code for our paper "Inverse Rendering of Near-Field mmWave MIMO Radar for Material Reconstruction".

Clone repository with submodules using git clone --recursive, or initialize after cloning using git submodule update --init --recursive.

We use Miniconda to install dependencies and require a NVIDIA OptiX >= v8.0 capable GPU and driver.

conda env create --file environment.yml
conda activate inv-radar

See MAROON (also included as a submodule), or download the example MAROON Mini Dataset, and extract to a folder of choice.

To execute differentiable radar rendering with default parameters on any dataset in MAROON, use:

python3 main.py /path/to/maroon/33_s2_hand_open/30

To use a different dataset, simply replace the respective part of the path argument, i.e. 33_s2_hand_open/30, accordingly. For example after extracting the MAROON Mini Dataset to data/maroon_mini/:

python3 main.py data/maroon_mini/02_cardboard/30

Each optimization run can be examined via Tensorboard, or by looking at the respective output in runs/, where a folder is created for each run using the following naming scheme: runs/<datetime>_<hostname>-<dataset>-<hash>.

Exemplary results for 02_cardboard/30 in layout (depth, normals, prediction, target, error map) using default parameters:

Exemplary results for 33_s2_hand_open/30 in layout (depth, normals, prediction, target, error map) using default parameters:

Run with different loss functions, as in Figure 9:

python3 main.py /path/to/maroon/33_s2_hand_open/30 --loss [l1, l1_complex, l2, l2_reco]

Run with different material regularization (storage) options, as in Figure 10:

python3 main.py /path/to/maroon/33_s2_hand_open/30 --material_storage [global, voxelgrid, hashgrid, vertex]

Run with different material models, as in Figure 11:

python3 main.py /path/to/maroon/33_s2_hand_open/30 --material [0-4]

Run with different features turned on or off, as in Figure 12:

python3 main.py /path/to/maroon/33_s2_hand_open/30 [--no_emptyfiltered, --no_reg_offset, --use_normalmap]

See main.py for a list of all possible command line arguments. Note that the --use_apc option will raise an exception per default, since this requires additional information regarding the antenna radiation pattern, which is not publicly available.

The authors would like to thank the Rohde & Schwarz GmbH & Co. KG (Munich, Germany) for providing the radar imaging device and technical support which made this work possible. This work was (partly) funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – SFB 1483 – Project-ID 442419336, EmpkinS. The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the NHR project b201dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the DFG – Project-ID 440719683.

@article{hofmann2025inverse,
    author={Hofmann, Nikolai and Wirth, Vanessa and Bräunig, Johanna and Ullmann, Ingrid and Vossiek, Martin and Weyrich, Tim and Stamminger, Marc},
    journal={IEEE Journal of Microwaves}, 
    title={Inverse Rendering of Near-Field mmWave MIMO Radar for Material Reconstruction}, 
    year={2025},
    pages={1-17},
    keywords={Backpropagation;MIMO radar;radar simulation;ray tracing;scattering parameters},
    doi={10.1109/JMW.2025.3535077}
}