May 11, 2018
Jolley Hall, Room 309
"Towards a Unified Framework of Structured Reconstruction of Indoor Scenes"
Adivsers: Tao Ju and Yasutaka Furukawa
We tackle the problem of reconstructing 3D structured models for indoor scenes. The challenge lies in the high-level reasoning of geometric structures in indoor scenes. We explore two properties of indoor scenes to address the challenging problem: 1) plane is representative of most indoor structures, and 2) structural surfaces (e.g., walls and floor) can be represented by a 2D floorplan. To exploit planarity, we propose a deep neural network to extract all plane structures from a static image as well as a novel approach to infer occluded geometries via layer decomposition of the scene. On the floorplan side, we address the problem of floorplan reconstruction from either a rasterized floorplan image or an image sequence with depth information. By combining both lines of research, we are pursuing a unified framework for structured indoor scene reconstruction with a mobile device. The reconstructed 3D model will enable many applications in augmented reality and indoor navigation.