Repository logo

Perceptual models for high-refresh-rate rendering



Change log


Dénes, György 


Rendering realistic images requires substantial computational power. With new high-refresh-rate displays as well as the renaissance of virtual reality (VR) and augmented reality (AR), one cannot expect that GPU performance will scale fast enough to meet the requirements of immersive photo-realistic rendering with current rendering techniques.

In this dissertation, I follow the dual of the well-known computer vision approach: vision is inverse graphics: to improve graphical algorithms, I consider the operation of the human visual system. I propose to model and exploit the limitations of the visual system in the context of novel high-refresh-rate displays; specifically, I focus on spatio-temporal perception, a topic that has received remarkably less attention than spatial-only perception so far.

I present three main contributions. First, I demonstrate the validity of the perceptual approach by presenting a conceptually simple rendering technique motivated by our eyes' limited sensitivity to high spatio-temporal change which reduces the rendering load and transmission requirement of current-generation VR headsets without introducing perceivable visual artefacts. Second, I present two visual models related to motion perception: (a) a metric for detecting flicker; and (b) a comprehensive visual model to predict perceived motion quality on monitors with arbitrary refresh rates and monitor resolutions. Third, I propose an adaptive rendering algorithm that utilises the proposed models. All algorithms operate on physical colorimetric units (instead of display-referenced pixel values), for which I provide the appropriate display measurements and models. All proposed algorithms and visual models are calibrated and validated with psychophysical experiments.





Mantiuk, Rafal


Computer Graphics, Perception, Rendering


Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge
Engineering and Physical Sciences Research Council (1778303)