AI-Accelerated Denoising in Sensor Simulation: Benefits of Adding Ground Truth

The utilization of ray tracing for sensor simulation enables the generation of bias-free simulation results, such as camera images, with a high level of physical correctness. However, due to the inherent Monte Carlo approach, ray tracing outcomes consistently exhibit simulation noise, especially at low iteration counts. This noise necessitates removal through additional denoising algorithms. In this context, the NVIDIA OptiX AI-Accelerated Denoiser effectively mitigates noise, facilitating the generation of more realistic images in less computing time, albeit at the potential expense of lost details during denoising, particularly when using a small number of iterations per image.
The presented video highlights the advantages of incorporating ground truth data (e.g., albedo and depth) as input to the AI. With this supplementary information, the AI denoiser can make more informed decisions to preserve details. Retaining more details while using fewer iterations in the simulation enables virtual validation of feature tracking algorithms (e.g., online calibration) at high simulation speeds.
www.rif-ev.de
www.mmi.rwth-a...
www.ficosa.com/
developer.nvid...