Depth Estimation with Manhattan World Cues on a Monocular Image

摘要

Many technologies requiring three-dimensional (3D) information, such as augmented reality, virtual reality, robots, and autonomous vehicles, are developing rapidly. Among them, 3D reconstruction from a monocular image is challenging because of the lack of information in monocular images and the large-scale errors from wide, low-texture regions found in them. However, a modern image obtained outdoors contains many artificial environmental stereo cues; hence, we regard this as a Manhattan world. In this study, we find vanishing points, which are effective stereo cues in the Manhattan world, to reconstruct 3D information from a monocular image. Then, 3D geometric information in this image is derived based on the vanishing points and relative positions and directions of straight lines. Hence, an initial depth map is generated accordingly. Depending on the 3D geometric information of each stereo cue found, accurate depth values are estimated and assigned to depth maps. As a result, the accuracy of depth information from an outdoor image obtained with the proposed method is improved by 37%, compared to existing algorithms.