Scientific
Publications
Details and downloads publications, technical reports and thesis papers.
Matting Research Projects
Temporal-Consistent Stereo Matting for High-Quality Novel View Synthesis and Visual Effects (2009)
This project develops new image processing techniques for the emerging field of 3D television. Given two videos that are recorded by slightly displaced cameras, we aim to extract (1) the opacity values of individual pixels, (2) a depth reconstruction of the scene and (3) the temporal relationship between the images of a sequence. The computed information can serve as the input for an autostereoscopic display that exploits our results to provide the user with temporal-smooth depth impression. Moreover, our results offer the exciting possibility for free-viewpoint video where the user has control over a virtual camera and can select the preferred viewpoint. To suit the needs of these applications, our focus lies on generating results of high-quality that outperform the current state-of-the-art. Due to addressing the image matting problem, we will also be able to handle fuzzy and hairy objects that are traditionally difficult in image processing.
Alpha Matting from Single and Multiple Images (2006)
This project aims to make a step forward in image matting, by developing novel algorithms and user interaction techniques that meet the requirements to quickly extract high quality objects from natural images. This research project is funded by Microsoft Research through its PhD Scholarship Programme.
In our "High-resolution matting framework (CVPR 2008)", we developed a new interactive approach for single image matting which splits the task of extracting a foreground object from a single background into two steps: Interactive trimap extraction and trimap-based alpha matting. By doing this we gain considerably in terms of speed and quality and in contrast to previous work are able to deal with high resolution images.
A number of matting algorithms rely on modeling the color of the user marked foreground and background regions to infer the optimal alpha value for every pixel. In our "Improved color model matting" paper (BMVC 2008), we exploit information from global color models to find better local estimates for the true fore- and background colors to finally estimate better alpha mattes.
