View Reconstruction from Uncalibrated Cameras for Three-Dimensional Scenes.

Abstract : With the development of faster memory and graphics hardware, there has been increased interest in the development of virtual environments. Researchers have been investigating the representation of mathematically defined objects and more recently the subsequent reconstruction of a view of the scene. The result has led to virtual worlds consisting of mainly geometric objects. Our interest is to extend this idea to real 3-D objects and scenes without resorting to modeling. While modeling uses little storage, the resulting images depend on the number of model parameters and require heavy computation. An alternative approach would be to use a very large number of views of the scene. This solution provides high quality images, but it also requires a large amount of memory. Hence, there is a tradeoff between storage requirements and accuracy of the reconstructed views. In this thesis, we propose an approach which offers high quality reconstructions with relatively low storage requirements. Our approach consists of a camcorder scanning the desired stationary object along several pre-specified trajectories. Neither the camera's exact motion nor its internal parameters are assumed to be known a priori. For certain locations in each trajectory, depth information is recovered using an adaptive region matching algorithm. The depth and corresponding intensity information are then used to generate accurate reconstructions. Results for intermediate views are excellent and preliminary results are promising for views not originally scanned by the camcorder. The eventual goal is to devise an environment consisting of more realistic and complicated objects. One possible system includes special hardware such as a stereoscopic display and a head tracking device. The viewer is able to gain the sense of 3-D by controlling the view of a scene by the movement of her head. We expect such a system with real-time reconstruction to become realizable in the near future.