G - O - V : Tutorial
G - O - V : Tutorial |
|
|
[an error occurred while processing this directive] | [an error occurred while processing this directive] |
Thessaloniki, Greece
In recent years, image-based rendering (IBR) techniques have advanced from static to time-varying scenes. Instead of still images, multiple synchronized video streams now capture a dynamic scene from different viewpoints. Suitably processed, a handful of video cameras during acquisition and a conventional PC for display is all that is needed to seat couch potatoes into the movie director's chair: video-based rendering techniques let the user interactively navigate his or her viewpoint through the real-world, dynamic scene to experience movie action or sports events from any self-defined perspective.
This tutorial explains in detail two different approaches how multi-video recordings may be processed in order to interactively render arbitrary new views of the recorded scene at high fidelity. The first part of the tutorial covers algorithmic as well as practical aspects of visual hull rendering. Visual hull reconstruction and rendering offers complete online processing capabilities and enables viewing the dynamic scene almost instantaneously. The tutorial explains different visual hull representations and their limitations. Joint reconstruction and rendering approaches exploiting programmable graphics hardware are described, and techniques are presented how to improve rendering quality and to minimize artifacts.
The second part of the tutorial is dedicated to an in-depth description of model-based multi- video analysis and free-viewpoint rendering. Here, a-priori information about scene content is exploited.Questions that will be answered are: How can a generic geometry model be automatically and robustly adapted to its video-recorded real-world counterpart using only silhouette information ? How can this offline technique be considerably accelerated by implementing a parallelized version of the algorithm ? How can texture information be additionally exploited to refine estimated body pose ? How must the video images be processed to avoid texturing artifacts due to geometry inaccuracies ? The complex movements of a human jazz dance performance illustrate the capabilities of the approach.