Three-dimensional models (3D) are enjoying wide applications in the movie and game industries. Over the past decade, image based modeling has emerged as a popular approach for producing realistic models from a set of photographs. Computer vision techniques have been developed that can automatically extract 3D surface representations from multiple images and have shown success in constructing high-resolution models, but are restricted to small-scale objects. On the other hand, many interactive modeling systems have been designed to allow users to define primitives on images for constructing 3D models via stereo triangulation, however these systems target man-made objects possessing geometric regularities and many technical issues have been ignored....[ Read more ]

Three-dimensional models (3D) are enjoying wide applications in the movie and game industries. Over the past decade, image based modeling has emerged as a popular approach for producing realistic models from a set of photographs. Computer vision techniques have been developed that can automatically extract 3D surface representations from multiple images and have shown success in constructing high-resolution models, but are restricted to small-scale objects. On the other hand, many interactive modeling systems have been designed to allow users to define primitives on images for constructing 3D models via stereo triangulation, however these systems target man-made objects possessing geometric regularities and many technical issues have been ignored.

This thesis presents a modeling framework that aims to integrate and take advantage of both automatic and interactive image based modeling approaches such that users can model large, real scenes comprising various kinds of objects. Contrary to existing interactive modeling systems, we first process the input image sequence with an automatic reconstruction algorithm to acquire a quasi-dense 3D point cloud used as input to our system. This additional 3D information provides the users with greater flexibility in geometric modeling and better assistance in texture extraction. We also demonstrate how this reconstructed 3D point cloud can be used for object removal, a functionality needed to handle occlusion in complex real scenes but absent in existing systems. Texture repairing tools have been developed and technical issues such as color inconsistency across images are addressed. We also discuss the difficulties in modeling a complete real scene due to its complex nature.