CD-Laboratory for Semantic 3D Computer Vision

The Christian Doppler Laboratory on Semantic 3D Computer Vision is concerned with the development of methods and algorithms, bridging the large gap between Machine Learning and geometric Computer Vision problems. Novel statistical algorithms are developed, that can be used as fundamental bricks in different 3D Computer Vision applications. Most solutions proposed are based on similar formalisms, even for very different problems. Moreover, special care is given not to sacrifice speed of the proposed methods, which is important for interactive applications and reliability. The techniques developed in the CDL make applications of 3D Computer Vision robust enough to work in realistic, uncontrolled conditions, even for non-experts and thus drastically extend their range of applicability. In particular, this addresses problems the industrial partner Qualcomm Technologies, Inc. is facing to move Robotics and Augmented Reality applications out of the lab, and cannot be solved by engineering only. Machine Learning methods also easily adapt to different types of input, as long as training data is available. Research in the CDL is not restricted to regular cameras and color images, but also considers other imaging sensors: Depth cameras are an obvious choice, but also infrared cameras and other very recent sensors including light-field cameras, which capture images from multiple viewpoints, and Dynamic Vision Sensors, which capture very fast motions, are considered. These sensors provide rich information complementary to the information provided by a regular camera, and therefore often multiple sources of input are used in the proposed approaches. 3D Computer Vision and Machine Learning are the topics of different, almost disjoint sub-communities in Computer Vision, however, the CDL is working right on the intersection of these areas.