Please note: This master’s thesis presentation will take place online.
Xiaochun Tong, Master’s candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Toshiya Hachisuka
Photorealistic image synthesis through physically based rendering is essential to achieve high visual fidelity. Monte Carlo rendering provides a scalable solution to accurately simulate the complex interactions between light and geometries within a scene. Monte Carlo integration numerically integrates a function by taking multiple pointwise estimations of the function and computing the average. Therefore, a typical Monte Carlo rendering algorithm independently samples one light path or path tree at a time for each pixel in the image. One approach to significantly reduce computational costs and enhance efficiency is to reuse light paths across multiple pixels and therefore amortize sample generation. The current state of the art of path reusing methods employs a technique known as shift mapping to deterministically map light paths from one pixel to another at low cost, yet the total computation cost is still linear w.r.t to the number of pixels processed in shift mapping.
We proposed a general framework, that generalizes reusing light paths to multiple pixels arranged in arbitrary two-dimensional shapes as image space shape splatting. Our shape is defined as a set of multiple pixels, and the framework allows us to sample such shapes more efficiently than by evaluating each pixel individually through shift mapping. Our key insight is to design a fast biased estimator that sparsely evaluates the contribution of shifted paths at random pixels within the shape and interpolates the contribution to the other pixels. We then employ a debiasing estimator to eliminate the bias from approximation. Our method can be seamlessly integrated with many state-of-the-art rendering methods and improving their efficiency over traditional pointwise sampling.