Depth estimation from a single outdoor image based on scene classification

Abstract

Recovering 3D depth from a single outdoor image is a basic problem in Computer Vision and Close-Range Photogrammetry. In this paper, an efficient depth estimation approach from a single outdoor image is presented. According to scene classification, depth of regions marked as sky, ground and vertical labels is respectively predicated. Firstly, a more accurate depth calculation model for ground regions is deduced from the camera imaging model, which takes camera pitch into account. Then the depth of vertical regions is estimated based on the ground regions depth. To consider the depth change of the regions, the vertical regions with visible ground-contain points are regarded as composed of one or multi-planes, and the vanishing points are incorporated into those points to further improve the position accuracy of intersecting lines. The contrastive experimental results are shown to verify the effectiveness and validity of the proposed approach.