FOCAS: Practical Video Super Resolution using Foveated Rendering

Figure 1: Left: low-resolution (LR) input; Middle: FOCAS takes 24 ms to construct a foveated high-resolution (HR) frame; Right: Traditional (non-foveated) super-resolution (SR) takes 64ms to construct a full HR frame.

Traditional SR upgrades the entire frame of the LR input to HR, incurring large computational costs, high latencies, and low frame rates, making it unsuitable for real-time video streaming. However, human vision is more sensitive to video quality in the central foveal region, and less sensitive in the periphery. FOCAS upgrades only the central foveal region of each frame to HR, resulting in reduced computational cost, lower latency, and higher frame rates, while nearly matching the perceptual video quality of traditional SR.