The focus of this thesis is to accelerate the synthesis of physically accurate images using computers. Such images are generated by simulating how light flows in the scene using unbiased Monte Carlo algorithms. To date, the efficiency of these algorithms has been too low for real-time rendering of error-free images. This limits the applicability of physically accurate image synthesis in interactive contexts, such as pre-visualization or video games.
We focus on the well-known Instant Radiosity algorithm by Keller [1997], that approximates the indirect light field using virtual point lights (VPLs). This approximation is unbiased and has the characteristic that the error is spread out over large areas in the image. This low-frequency noise manifests as an unwanted 'flickering' effect in image sequences if not kept temporally coherent. Currently, the limited VPL budget imposed by running the algorithm at interactive rates results in images which may noticeably differ from the ground-truth.
We introduce two new algorithms that alleviate these issues. The first, clustered hierarchical importance sampling, reduces the overall error by increasing the VPL budget without incurring a significant performance cost. It uses an unbiased Monte Carlo estimator to estimate the sensor response caused by all VPLs. We reduce the variance of this estimator with an efficient hierarchical importance sampling method. The second, sequential Monte Carlo Instant Radiosity, generates the VPLs using heuristic sampling and employs non-parametric density estimation to resolve their probability densities. As a result the algorithm is able to reduce the number of VPLs that move between frames, while also placing them in regions where they bring light to the image. This increases the quality of the individual frames while keeping the noise temporally coherent — and less noticeable — between frames.
When combined, the two algorithms form a rendering system that performs favourably against traditional path tracing methods, both in terms of performance and quality. Unlike prior VPL-based methods, our system does not suffer from the objectionable lack of temporal coherence in highly occluded scenes.
