Image-Based Modelling and Rendering of Humans

Prof. Dr. Peter Eisert, Humboldt-Universität Berlin/Fraunhofer-Institut für Nachrichtentechnik

Date: 19/04/2013
Time: 14:15-15:15
Room: H-F 114

In recent years, digital cameras have significantly increased in quality and resolution, while dropping prices make them available for many different applications. Current high resolution images now show many details of a scene that are otherwise difficult to capture, e.g. with scanners or other 3D sensors. In this talk, methods are presented that exploit multiple high resolution views for image-based modelling and rendering of 3D scenes. A method for high-detail 3D geometry estimation is shown that exploits the visible details of material structures, like the skin for face reconstruction. While such approaches lead to detailed 3D models, rendering of 3D scenes can also be conducted directly on the images. We present a method for image-based rendering of dynamic objects like humans, that interpolates new views from an image database with different poses. Only approximate geometry is required to allow for animation while surface details and shading effects are represented by the appearance. This is illustrated with results for the particular application of virtual clothing, where pieces of clothing are animated and rendered by interpolation from a set of offline captured images.