Manzi, Marco (2016). Advanced Techniques in Gradient-Domain Rendering (Unpublished). (Dissertation, Universität Bern, Philosophisch-naturwissenschaftliche Fakultät)
PhD Dissertation Marco Manzi.pdf - Accepted Version
Available under License BORIS Standard License.
Download (68MB) | Preview
Rendering realistic images requires solving the notoriously hard physically-based light transport problem. Almost all of the state-of-the-art physically-based rendering methods use Monte-Carlo sampling of the light paths contributing to the image. These methods suffer from variance until convergence. Depending on the scene, an impracticable amount of time might be required to get a clean image. A recently developed method called gradient-domain Metropolis light transport mitigates this problem. It first samples the image-space finite differences of paths alongside the paths themselves and then reconstructs a clean image from the sampled data by applying a screened Poisson reconstruction. The method exploits two properties: first, the gradients of natural images are usually much sparser than the image itself, thus sampling efforts can be concentrated in fewer regions of the path space. Second, sampling finite differences allows using correlated sampling in rendering, which can strongly reduce noise in the finite differences. Both properties combined lead to dramatical speed-ups compared to classical (Markov-chain) Monte-Carlo rendering methods. This dissertation builds up on the aforementioned gradient-domain Metropolis light transport and proposes a number of improvements and generalizations. We improve the sampling by replacing finite differences by arbitrary differences and by combining different sampling strategies in an unbiased way. We also generalize the method to non-MLT rendering methods like bidirectional path tracing. Further, we develop an algorithm that regularizes the screened Poisson reconstruction by using auxiliary scene information in order to increase image quality. This leads to the first method that combines gradient-domain rendering with classical image-space denoising. And finally, we incorporate temporal finite differences in gradient-domain rendering in order to create stable animations, thus making gradient-domain rendering an even more appealing option for production rendering.
|Item Type:||Thesis (Dissertation)|
|Division/Institute:||08 Faculty of Science > Institute of Computer Science (INF) > Computer Graphics Group (CGG)
08 Faculty of Science > Institute of Computer Science (INF)
|UniBE Contributor:||Manzi, Marco|
|Subjects:||000 Computer science, knowledge & systems|
|Date Deposited:||29 Dec 2016 13:15|
|Last Modified:||29 Dec 2016 13:15|