Presenter

Mr Beerend Ceulemans - Vrije Universiteit Brussel [Email]

Abstract

With the increasing interest in high-quality 3D multimedia content over the last decade, a great deal of research is dedicated to further improvement of already existing technologies as well as the search for new ways to present 3D multimedia content to consumers. To this end, significant improvements have been made in the area of stereoscopic 3D video which is currently the most ubiquitous type of 3D. With the release of the movie Avatar in 2009 it was clear that stereoscopic 3D was within our reach and both industry and academia have made various contributions in the acquisition, storage, transmission, and visualization of such content.
In contrast, others were not fully satisfied with the fact that stereoscopic 3D requires consumers to wear appropriate glasses in order to perceive depth differences. To this end, several types of visualization technologies were developed that can simultaneously display multiple different viewpoints for different viewing positions. These so-called autostereoscopic devices have the advantages that they can deliver content to multiple users without the need for special glasses. Additionally, these devices also offer different viewpoints to a moving user hereby enabling the motion parallax effect which is an important depth-cue for the human visual system but which is not present in classical stereoscopic 3D. The downside of existing autostereoscopic displays is that the current state-of-the-art in acquisition and processing is not sufficiently mature to push this technology to consumers.
This thesis presents a new color correction method that has been developed to handle color differences between cameras, using a global energy minimization method to provide an appropriate color transfer while preserving important structural information. A second contribution is the development of a multiview synthesis framework that is able to render novel viewpoints from a sparse set of captured texture and depth pairs. A point cloud-based filter eliminates unreliable measurements while careful edge processing avoids ghosting artefacts. Furthermore, when generating novel views, parts of the scene that were initially occluded for the capturing sensors may become visible in the synthesized views, corresponding to the so-called disocclusions. The proposed framework includes an inpainting method to fill-in these disoccluded regions by minimizing the energy of a Markov Random Field. As case study, we employ our methods to generate dense multiview content for an autostereoscopic 3D display that is present in the lab. The proposed methods are key enabling technologies for the future wide-scale deployment of autostereoscopic 3D video.

Short CV

MSc Applied Sciences and Engineering: Computer Science, VUB, 2013