Fotogerçekçi imaj sentezleme için uyarlanabilir yöntemler

Efficient sampling in light transport simulation algorithms used for photorealistic image synthesis remains an unsolved problem. In this thesis, we first explain the basics of light transport simulation and main challenges that prevent efficient sampling. We emphasize adaptive importance sampling, which is a promising approach to overcome these challenges. Thus, we review previously published light path guiding methods that enable adaptive importance sampling. We state that most of these guiding methods only approximate incident radiance at points of scene's surfaces and generate samples according to this approximation. This approach works relatively well for diffuse surfaces. However, samples generated in this way at points on glossy surfaces may become invalid, since only a subset of incident light is reflected as glossy materials are view-dependent. In this thesis, we propose a simple extension to a previous tree-based guiding method to make it aware of view-dependent light transport. As in the previous work, spatial domain of the scene is adaptively partitioned into regions. To each of these regions, a quadtree that records incident radiance is attached. In our method these quadtrees are used only for diffuse surfaces. We attach to each spatial region an additional quadtree that partitions the space of view directions and stores quadtrees that record incident radiance. These quadtrees are then used for glossy surfaces. With our extension, we observe improvements in scenes with many glossy surfaces and minimal memory overhead in scenes without view-dependent light transport.