16:00 - 17:30 - May 20 (Thursday)
Session chair: Ioannis Stamos, City University of New York
The paper proposes an automatic method for fast and accurate alignment and multi-view segmentation of city-scale street view data. Simple 3D building outlines are used to bootstrap and automate all three steps of the algorithm: alignment with LIDAR data, image segmentation using graph cuts and its multi-view refinement. The multi-view segmentation produces consistent labels for each building in multiple images. The method can process city-scale datasets in approximately a day on a single server.
To capture fast-moving objects with active 3D scanning systems, a oneshot scanning method using a single static image is preferable than using multiple patterns. Recently, oneshot scanning methods that use intersection points of grid patterns have been proposed. In those methods, solutions only from intersection points have ambiguity, and using additional information such as variation of grid intervals was crucial. However, it makes the method difficult to increase the density of the pattern. In this work, by using multiple projectors, each of which projects parallel line patterns instead of grid patterns, shape is reconstructed using the intersection points between patterns. It is shown that, by using two projectors, a unique linear solution is possible for such a system, thus, dense grid patterns with uniform intervals can be used. Furthermore, blind areas caused by occlusion and self-occlusion are drastically reduced.
This paper proposes a practical system for high-resolution 3D acquisition of facial performance using standard video and lighting equipment. The system combines multiple view and photometric stereo with visual marker motion capture to acquire high-detail 3D models which are temporally registered. The photometric stereo based on passive colour illumination reconstructs fine detail of skin wrinkles and pores on the pixel resolution at every frame. In comparison to existing facial capture systems which require high-speed illumination and video recording, standard video-rate cameras combined with three colour lights provide similar quality of resulting 3D model.
The rolling shutter acquisition mode of CMOS cameras, which sequentially exposes the scan-lines, creates distortions in the image when the single filmed object moves with respect to the camera. Previous work showed that this can be used to estimate the initial object pose and time-constant kinematics. We propose a generic model for rolling shutter cameras capable of dealing with both uniform and non-uniform motion, contrary to previous work. Its application to the estimation of a dynamic pose is validated using experimental results on both synthetic and real data. The proposed framework outperforms previous work when dealing with non-uniform motion.
