Do You See the Shape? Diffusion Models for Noisy Radar Scattering Problems

Determining the shape of 3D objects from high-frequency radar signals is analytically complex but critical for applications like aerospace and autonomous driving. Previous methods using deep learning have successfully been applied to this task, but the radar response in practical applications contains noise, which is hard to model deterministically. In addition, we often only observe the 3D object from partial viewing angles, leading to a complex one-to-many mapping task. In this work, we demonstrate that diffusion models are a suitable learning paradigm for radar inverse modeling due to their probabilistic learning and denoising properties. We present the radar2Shape model, which approximates the distribution of shape parameters conditioned on radar responses that are representative of practical applications. In addition to being more accurate than a deterministic competitive baseline across levels of noise, we show that the probabilistic nature of radar2Shape is important to capture the uncertainty associated with object reconstruction with partial data.