Single image-based 3D scene estimation from semantic prior

Abstract

Reconstructing a three-dimensional (3D) structure from a single image sequence to provide relevant contextual information for better human visual perception is a fundamental problem in computer vision. A 3D scene estimation methodology from a segmented image sequence that is learned from semantic priors is proposed. In particular, semantic information including 3D geometric characteristics can very efficiently predict the 3D structure of the scene from a given semantic region. The approach, which utilises semantic priors to estimate a 3D scene, is very robust for direct 3D scene reconstruction from an ambiguous depth map. The efficiency and effectiveness of the proposed approach has been proven experimentally with a large database.