An Experimental Evaluation of Debayering Algorithms on GPUs for Recording Panoramic Video in Real-Time

2. System Overview

We have earlier described our panorama pipeline used in the Bagadus system [Halvorsen et al., 2013, Stensland et al., 2014]. In this system, we record raw 2040 × 1080 video frames in Bayer format from five cameras at 50 fps, and each of these camera streams must be debayered in real-time.

Modern GPUs can provide significantly better performance than a CPU for certain tasks. They are optimized for applying small transformations to every single pixel or texture element, with hundreds or thousands of threads performing the same task in parallel, with minimal inter-thread communication. Debayering is an inherently parallel operation, as each pixel, or block of pixels, can typically be calculated locally. Hence, with our high data rate and real-time requirements, we found the GPU to be better suited to perform this task. However, more complex algorithms that require a greater level of contextual information about the entire image will not achieve the same performance increase.