Our work at the Computer Vision lab is primarily concerned with geometric estimation problems in 3D vision. Conventional approaches to solving for the 3D scene (structure) and cameras (motion) involve expensive non-linear optimizations. We have developed a unified geometric framework known as motion averaging that efficiently solves the camera geometry problem inherent to structure-from-motion. Our approach poses the problem as one of estimating individual camera motions given a number of pairwise relationships. We extensively exploit the rich geometric structure of finite-dimensional Lie groups to develop efficient, scalable, accurate and robust solutions. This motion averaging approach yields state-of-the-art accuracy for large-scale problems and is commonly used in many 3D reconstruction pipelines.
Another area of recent interest has been recovering high accuracy 3D scans of objects using depth cameras. While depth camera scans are noisy, photometric normals yield high-quality local information for 3D surfaces. We have developed a unified framework for accurate depth-normal fusion that results in very accurate 3D scans of objects with fine-scale 3D detail.
Much of our current work is focused on analyzing the implications of our motion averaging framework for SLAM problems in robotics and visual navigation.