The British Machine Vision Association and Society for Pattern Recognition 

BibTeX entry

  AUTHOR={Spela Ivekovic},
  TITLE={Disparity Map Completion for Trilinear-Tensor View
    Synthesis from Wide-Baseline Stereo},
  SCHOOL={Heriot-Watt University},


In this thesis, we present a model-based disparity-map completion method for high-quality novel-view synthesis from wide-baseline stereo, using trilinear-tensor transfer. In immersive videoconferencing environments, the wide-baseline stereo setups are used to gain maximum information about the scene with a minimum number of cameras in order to limit the amount of data that must be processed or transferred. To promote the immersive experience, the cameras are arranged around the screen with their viewpoint differing from that of a remote videoconference participant. The correct viewpoint is synthesised by means of novel-view synthesis. When using trilinear tensor to synthesise novel views, dense disparity maps between the stereo pair of images are crucial to the quality of the synthesised view. However, as the scene in a wide-baseline stereo setup is imaged from two relatively different viewpoints, the correspondence search presents a difficult problem which does not result in dense disparity maps. In order to complete them, some form of post-processing is normally used. As a simple interpolation of the available disparities produces limited results, we look at using prior knowledge about the scene in the form of a human body model to complete the missing disparities for high-quality view synthesis in a videoconferencing environment. We describe two alternative methods to disparity completion with a generic body model, the first working in 3-D space and the second in disparity space. We present an articulated subdivision surface model of a human body in 3-D and its equivalent in disparity space, address articulated pose estimation in 3-D and disparity space using particle swarm optimisation and present a quasi-interpolation method for fitting a subdivision surface body model to an unstructured, incomplete and noisy cloud of data points in 3-D and disparity space. Finally, we show the results of novel-view synthesis with disparity maps completed using the generic body model.