Real-Time 3D Urban Scene Reconstruction from Video

Recent developments in computer vision have improved the performance of structure from motion techniques for 3D model reconstruction from video. The current research challenge is to achieve reconstructions at interactive frame rates and to enlarge the numbers of images used from a few thousand frames at most to several millions of frames. This opens a wide field of applications for these techniques as well as it boosts the research in computer vision to improve the reliability of the methods. The talk focusses on the work on realtime methods for reliable automatic geo-registered large scale 3D scene-reconstructions from millions of images for a range of capture systems from a sensor-augmented multicamera system to a single hand-held video camera. Besides high quality in terms of both geometry and appearance, the used techniques deliver real-time performance deploying CPU and GPU.