Home | Search | Help  
Home Page Università di Genova

Seminar Details

Date 30-10-2008
Time 14:30
Room/Location Sala conferenze del DISI (aula 322, 3 piano)
Title Novel-view synthesis for immersive videoconferencing
Speaker Spela Ivekovic
Affiliation School of Computing, University of Dundee, UK
Link https://www.disi.unige.it/index.php?eventsandseminars/seminars
Abstract The immersive videoconferencing station consists of a large plasma display surrounded by cameras recording the person sitting in front of the screen. The remote conference participants are shown on the display. The cameras see the local person from a different viewpoint than the one expected by the remote conference participants. Novel-view synthesis is used to adjust the viewpoint to better match a natural conversation scenario. As the bandwidth and computation time are limited, only a relatively small number of cameras can be used at each videoconferencing station. Consequently, at each station, the cameras must be arranged as wide-baseline stereo setups and the disparity maps generated from such stereo pairs, and required for high-quality novel-view synthesis, are noisy and incomplete. In order to improve the view synthesis quality, the disparity maps are usually post-processed with an interpolation algorithm, using a minimal amount of prior knowledge about the scene. In this talk, I will discuss an alternative solution, where a generic human body model is used to complete the disparity data. Within this context, I will present the articulated body pose estimation approach with particle swarm optimisation, fitting the human body model to noisy disparity data with a quasi-interpolation technique and synthesising novel views with the resulting complete model-based disparity maps. While all these steps can be performed in 3-D, I will also present an alternative approach which completes the disparity map in disparity space to avoid error propagation and reduce computational effort.
Back to Seminars