Sub-Scene Level Analysis and Synthesis of 3D Indoor Scenes

3D indoor scenes are ubiquitous in computer graphics applications such as 3D games and interior design. With the emerging applications in VR/AR, there is an increasing demand of realistic 3D scene data. However, designing 3D indoor scenes requires proficient 3D modeling skills and is often time-consuming. A promising solution to the content-creation bottleneck of scenes is to utilize the existing scene data for data-driven 3D scene generation. Recent research about data-driven indoor scene processing in computer graphics usually takes a holistic view and operates at the object-level for scene analysis and synthesis. The main limitation of existing methods is their applicability to characterizing and modeling complex scenes. In this thesis, we address the problems of data-driven 3D indoor scene analysis and synthesis via sub-scene level processing. Our goal is to improve the understanding of scene structures through sub-scene level analysis and develop efficient systems to create complex scenes by manipulating sub-scenes instead of individual objects.To this end, we first introduce focal points, the representative sub-scenes, for characterizing, comparing, and organizing collections of complex and heterogeneous data, and apply the developed concepts and algorithms to collections of 3D indoor scenes. Then, we propose a framework for action-driven evolution of 3D indoor scenes. Human actions learned from annotated photographs are applied to trigger appropriate object placements at a sub-scene level, inducing a more compact way of scene generation. Finally, we present a novel framework that uses natural language to generate 3D indoor scenes. We demonstrate advantages of focal-centric scene comparison and organization over existing approaches. We show results of our action-driven and language-driven scene synthesis that lead to realistic, messy and complex 3D scenes, and evaluate the plausibility and naturalness of the scenes by user studies.