Research

Computer vision has entered a new and exciting phase: many of its sub-problems have reached a stage, where they are becoming applicable to real-world scenes. Present solutions to classical tasks like segmentation, object recognition, tracking, and ego-motion estimation deliver usable (although far from perfect) visual information about the world. We think that in the next phase it will become increasingly important to combine several of these visual information channels, rather than incrementally advancing a single one. Our hope is that through the integration it will become possible to solve more complex problems and build more powerful systems.

The joint solution of two or more visual tasks runs as a common thread through our research. We aim to explore the connections between different channels of visual information, and the mutual benefits of concurrently solving tasks pertinent to the same final goal. Channels are often complementary and support each other, and in close connection can solve harder problems than any single one could. Moreover, exciting connections beyond the classical domain of computer vision have emerged to machine learning, cognitive science, and computer graphics, calling for further integration in a more comprehensive approach to "visual information processing".

Two examples of recent work are (1) an integrated approach to object detection, tracking, and visual localization, in order to enable autonomous navigation and driver assistance in urban scenarios, and (2) joint segmentation and structure-and-motion estimation to allow 3D reconstruction of dynamic scenes (see also our publications page).