Differentiable rendering for synthetic aperture radar imagery




Wilmanski, Michael Charles

Journal Title

Journal ISSN

Volume Title



Deep Neural Networks (DNNs) have been at the forefront of modern advancements in computer vision. In recent years, there have been efforts to incorporate more domain knowledge into DNN training and architectures by inserting signal processing functions into optimization pipelines. This can lead to DNNs that are trained more robustly and with limited data, as well as the capability to solve challenging inverse problems. Consequently, there is increasing interest in differentiable rendering, which allows explicitly modeling geometric priors and constraints into an optimization pipeline. Existing efforts in differentiable rendering have focused on imagery from electro-optical sensors, particularly conventional RGB-imagery. In this report, I propose an approach for differentiable rendering of Synthetic Aperture Radar imagery and demonstrate proof-of-concept experiments for the proposed method on the inverse problem of 3D Object Reconstruction. I also discuss limitations of the approach and next steps for future work.


LCSH Subject Headings