Combining approximate geometry with view-dependent texture mapping - a hybrid approach to 3D video teleconferencing

We present a hybrid system using computer vision and graphics methods that effectively combines fast automatic 3D model extraction with view-dependent texture mapping for the purpose of real-time person-to-person 3D video teleconferencing. In our approach the problem of creating a dense disparity map is bypassed. We compute only a sparse geometric model, which is textured from multiple cameras using an image-based rendering approach. Using a commodity graphics card and three personal computers arranged in a cluster, we achieve a performance of 3 to 4 frames per second. We demonstrate the ability to provide a 3D tele-immersive environment with our system.