Torchbp

Fast C++ Pytorch extension for differentiable synthetic aperture radar image formation and autofocus library on GPU.

Only Nvidia GPUs are supported. Currently, only some of the operations are supported on CPU.

On RTX 3090 Ti backprojection on polar grid achieves 225 billion backprojections/s and fast factorized backprojection (ffbp) is ten times faster on moderate size images.

Installation

Tested with CUDA version 12.9.

From source

git clone https://github.com/Ttl/torchbp.git
cd torchbp
pip install --no-build-isolation -e .

Documentation

API documentation and examples can be built with sphinx.

pip install .[docs]
cd docs
make html

Open docs/build/html/index.html.