Single-image 3-D depth estimation for urban scenes

Abstract

In this paper, we focus on recovering a 3-D depth map from a single image. Given an image of urban scene, we extract linear perspective information to establish the 3-D scene model. Unlike approaches which use only occlusion relationship between objects to estimate the relative depth of the image, we further combine the perspective geometry information with the occlusion relationship between objects. Besides, we propose the construction of depth gradient maps to represent the depth variation trend along the vertical and horizontal directions. The image is first partitioned into geometric components and initial depth gradient maps are generated based on the relative position between the vanishing point and the classified components. Incorporating main directions of vanishing lines and occlusion boundaries in the initial depth gradient maps, a refined depth map is obtained by using a CRF (conditional random field) model. We demonstrate that our approach can produce realistic relative depth maps for images of urban scenes.