CAREER: Structured Indoor Modeling

Washington University, Saint Louis MO

Investigators

Abstract

Many objects around us are associated with certain functions. We use a switch to turn on a light, push a button to call an elevator, and open a door to enter a room. While 3D digital mapping is getting increasingly more attentions due to the rapid advancements in the 3D reconstruction techniques and the 3D sensing hardware, current methods are merely optimized for geometric fidelity and lack such functional information. This project discovers rules governing indoor scenes from a database of indoor 3D models, and then develops an algorithm to reconstruct functional indoor 3D models. High fidelity 3D models, if equipped with functional information, would facilitate fundamentally new applications that influence our lives at a much deeper level. The developed technologies have many different applications, from helping indoor navigation to assessing compliance with building codes, accessibility codes, and energy efficiency levels. The broader impacts of the project include architecture, civil engineering, urban geography and sociology, real estate, and transportation. The research of this project is well integrated with the education. The education plan includes an interdisciplinary course with Sam Fox School of Design & Visual Arts at Washington University, and co-developing a K-12 teaching module with a local high school teacher, which is estimated to impact 2,500 high school students. The central idea of this research is to discover structural elements constituting an indoor scene, together with their hierarchical and functional relationships. For example, a building consists of stories, each of which contains rooms, each of which contains windows. Doors connect rooms, and an elevator door is a passage connecting different floors. First, the project defines a novel structured scene representation together with its rigorous structure grammar. Second, the project derives a principled new structured reconstruction algorithm that follows the rules in the structure grammar. Third, the structured model representation and reconstruction algorithm open up new opportunities to enable a highly tunable reconstruction system. This project develops a system that is capable of enforcing richer classes of geometric constraints far beyond existing methods, and controlling the properties of an output model effectively to meet the demands of specific applications directly. The developed technologies can be transformed to other geometry reconstruction and shape extraction problems, such as 2D shape segmentation, 3D outdoor architectural modeling, and 4D dynamic scene reconstruction.