3D information extraction from monocular images

Abstract

Extracting depth information from stereo images has nowadays become common practice with many well known and established techniques. The problem is much harder when only a single monocular image is used in order lo approximate depth information. In this dissertation the problem of 3D information extraction from single monocular images is addressed. We present an automatic method for creating a simplified 3D model from a single monocular photograph using visual heuristics. The aim of these heuristics is to align a 3D template mesh, the shape of a frustum, to a picture. This effectively transforms the 2D image into a simplified 3D model using depth estimation. The image is processed through a three stage pipeline: statistics, template fitting and point cutting. Using statistics and information extracted from the image the position of the eight main points that correspond to the shape of a frustum is estimated. These points are converted into planes and a point cutter then creates textures from this information which are then mapped to the 3D mesh. The results are then displayed in an interactive environment where the user can navigate through the generated model. These methods are then evaluated on a set of images ranging from simple computer generated images to more complex real world examples. In general the results achieved indicate that the use of visual heuristics alone do not provide adequate information to accurately make depth estimations.