Systematic Evaluation of the Quality Benefits of Spatiotemporal Sample Reprojection in Real-Time Stereoscopic Path Tracing

Abstract

Path tracing is a commonly used but computationally highly expensive stochastic ray tracing method for rendering photorealistic visual content. Combined with a real-time constraint, for example in stereoscopic virtual/augmented reality applications, it typically limits us to rendering at most a few samples per pixel, yielding very noisy results. However, the spatial and temporal redundancies are commonly utilized by reprojecting existing samples between different viewpoints and frames, thus cheaply improving the quality. We provide new insights to the quality benefits of reprojection by systematically evaluating the effective quality of spatiotemporally reprojected stereoscopic path traced data. We show that spatiotemporal reprojection increases the quality of 1 sample per pixel (spp) data by almost a factor of 25 on average, in terms of the effective spp count of the result. Since we are able to reproject 94-98% of the samples, only the remaining 2-6% of the samples in the target frame need to be path traced. We also evaluate how the quality improvement gained through spatiotemporal reprojection scales as the number of input samples per pixel increases, showing that the highest gains are achieved at the lowest input spp counts. Finally, we show how blending existing path traced data and stereoscopically reprojected data further improves the quality of spatiotemporal reprojection, on average yielding a 47% higher effective spp than without blending.

Real-time Stereoscopic Image-parallel Path Tracing

“…Prior works related to spatial reprojection [3,14,15,8] are focusing on single GPU utilization. In this paper, we consider multi-GPU rendering and the relative distribution of the workload, including the real-time constraint as well.…”