Abstract
Deep generative models (DGMs) have been studied and developed primarily in the context of natural images and computer vision. This has spurred the development of (Bayesian) methods that use these generative models for inverse problems in image restoration, such as denoising, inpainting and super-resolution. In recent years, generative modelling for Bayesian inference on sensory data has also gained traction. Nevertheless, the direct application of generative modelling techniques initially designed for natural images on raw sensory data is not straightforward, requiring solutions that deal with high dynamic range signals (HDR) acquired from multiple sensors or arrays of sensors that interfere with each other, and that typically acquire data at a very high rate. Moreover, the exact physical data-generating process is often complex or unknown. As a consequence, approximate models are used, resulting in discrepancies between model predictions and observations that are non-Gaussian, in turn complicating the Bayesian inverse problem. Finally, sensor data are often used in real-time processing or decision-making systems, imposing stringent requirements on, e.g. latency and throughput. In this article, we discuss some of these challenges and offer approaches to address them, all in the context of high-rate real-time sensing applications in automotive radar and medical imaging.This article is part of the theme issue 'Generative modelling meets Bayesian inference: a new paradigm for inverse problems'.