Abstract:
This thesis addresses theoretical basics and applications of graphically interactive web- based image processing and computer vision frameworks which extract depth informa- tion from di erent forms of digital two-/three-dimensional (2D/3D) inputs and other col- laborative online research meta-tools of 3D data generation, visualisation, and analysis for stereo images. Such e cient and e ective online processing tools have not been ade- quately explored so far. The thesis also discusses various novel and existing algorithms and practicable functionalities of many techniques such as camera control, calibration and recti cation, 3D data segmentation, 3D graphic rendering, face modelling, feature detection with active contours, image matching, and stereo reconstruction. Within this thesis, a guided dynamic programming technique is proposed. It uses full or partial pro- les generated by other algorithms to guide the process an e cient symmetric dynamic programming stereo matching to produce higher-quality disparity maps. The evaluated results have shown that this algorithm was demonstrably accurate, fast, and relatively robust for stereo correspondence when tested with modern evaluation techniques and datasets. Along with the theory development, a versatile interactive system is built, allowing users to dynamically reconstruct visible surfaces of 3D scenes from stereo pairs of im- ages acquired with o -the-shelf monocular and stereo cameras including cell-phones and webcams. The system communicates with potential users via the Internet and allows uploading of individual static images or live video sequences. Images are automatically or manually co-aligned into epipolar stereo pairs using either calibrated or uncalibrated image recti cation. A recti ed pair is processed by one of the available stereo matching algorithms to reconstruct a visible surface of a 3D scene. The user's computer or cell- phone is used to provide an immediate, controllable, 3D display. Within a short waiting time, many results are returned to the user including aligned 3D side-by-side photos, dis- parity and texture maps, anaglyphs, autostereograms, virtual 3D scenes, 3D .OBJ les, or even live depth video sequences; all of which can be shared with others in a public gallery. This portable, simple to set up and operate system, currently very rare on the Internet, is located at http://www.ivs.auckland.ac.nz/quick_stereo and has been visited by thousands of people spread across more than 50 countries. At the present, the public gallery page contains 3,000 real life stereo datasets in over 30 categories; each of them has a pair of raw images, recti ed images, and other 3D related data, which are be- lieved to be useful for testing and learning purposes. A few further potential applications such as remote camera control for online calibration and recti cation, stereo-vision-based autostereogram creation, modelling of a 3D facial avatar, stand-alone stereo reconstruc- tion application for commodity stereo cameras, and mobile-phone applications are also discussed.