Sample Reprojection for Real-Time Light Field Path Tracing
2021, Laukkanen, L., Master of Science thesis, Tampere University.
Abstract: Path tracing can be used to render photorealistic images but requires high performance computation. Visualising virtual reality content with path tracing could offer immersive and photorealistic experience. For comfortable user experience, virtual reality devices and applications target for high framerates and display resolutions, which can be a huge computational workload even for conventional rasterization pipeline. Light field displays can eliminate the vergence-accommodation conflict, which rises from mismatch of visual cues from stereoscopic displays and viewed virtual content. The conflict can manifest itself as a motion sickness and nausea. Light field HMD could provide a solution to the vergence-accommodation conflict, but introduces a high computational workload, as light fields involve rendering multiple different viewpoints. Applying light fields with path tracing, achieving real-time interaction can be challenging, as the computational complexity is high.
Sample reprojection techniques can offer lightweight methods to increase the performance of path tracing by reusing already calculated pixels. Reusing pixels in image space is computationally less demanding compared to determining the pixel colour by path tracing. In this thesis, we consider a light field as a grid of multiple viewpoints. The number of path tracing calculations is affected by how the sources for reprojection are selected. These different sources are compared, how they affect the sample usage and which setups are best to be utilized to avoid unnecessary path trace calculations. Different sampling setups of source views are reprojected spatially and temporally to create a synthetic light field. How each sampling setup affects the quality and pixel invalidation scalability is also evaluated.
Results indicate, that sample reprojection can alleviate required calculations for multiview path tracing. At best, spatially 4.89 % and temporally 4.57 % of samples were used from the budget reserved to render all multiple viewpoints. Best sample reuse ratio of 3.22 % was reached by combination of both temporal and spatial sample reprojection. At worst case, spatially 12 % of samples were required from a total budget to render the light field. Reprojection could offer roughly 10 - 20 times reduction of required path traced samples for multiple views. Spatial reprojection offers best sample reuse when the samples are distributed near the edges of the light field multiview grid.