3D Scene Understanding from a Single Image

The main theme of this thesis is 3D scene understanding from a single image. We start with utilizing k-d trees to partition point clouds to capture both local and global structure, and continue with inferring complete point clouds from a single image via depth intermediation. Then, we propose a pipeline to jointly estimate the depth and 3D layout of an indoor scene from a single panorama image. We conclude with reconstructing the 3D indoor semantic scene point clouds from a single panorama image. Findings of this thesis may lead to a better understanding of 3D reconstruction from single images.