Visual Integrity in the Age of AI: An Evaluation of DLSS and DLAA in Geospatial Visualization

Keywords

Machine Learning Techniques; Computer Graphics Techniques; Guidelines; Geospatial Data; Terrain, Point Clouds; Deep Learning Super Sampling, DLSS, Upscaling; DLAA; Anti-Aliasing

Abstract

Deep Learning Super Sampling (DLSS) is a relatively new machine learning technology for accelerating 3D graphics, by rendering at lower resolutions, and inferring what the higher resolution output should look like. However, DLSS was designed for, and trained on, video games. Can this technology, intended for the entertainment market, be repurposed to accelerate high-resolution scientific and geospatial visualization? This paper evaluates the visual accuracy costs of inferring higher resolution data visualizations with DLSS upscaling. A related technology, Deep Learning Anti-aliasing (DLAA) is similarly evaluated. This evaluation focused on 3D geospatial applications, with (primarily non-photorealistic) terrain rendering and point cloud rendering, though the results are applicable to a wider range of scientific visualization data types. The results demonstrate that DLSS/DLAA can significantly impact visual accuracy, and should be avoided for visualizations requiring accurate portrayal of fine details.