Date: December 13, 2017
Estimating scene motion accurately and efficiently is a key to many applications such as robot planning, dynamic reconstruction, autonomous driving, action recognition, etc. In this report, I will present my research in dynamic scene motion understanding in 3D using images from the examples in dynamic object mapping and scene flow estimation. I will start with my work and motivation in dense scene mapping in the dynamic world, in both RGB-D indoor scenario and stereo outdoor large-scale scenes. Then I will focus on discussing my two recent work in dense 3D non-rigid correspondence, termed as scene flow. In the first work, I will present how to solve scene flow via continuous optimization method applied in stereo imagery. To overcome the inference complexity in optimization-based approach, I propose the first learning scene flow method applied in RGB-D images, which achieves state-of-art accuracy, fast execution time and good generalization to complex scenes.
Zhaoyang Lv is a Ph.D. student in Robotics at Georgia Tech, School of Interactive Computing, jointly advised by Prof. James Rehg, and Prof. Frank Dellaert. Before he starts the Phd at Georgia Tech, he finished the Master thesis under the supervision of Prof. Andrew Davison at Imperial College London. His research interests cover computer vision researches as the perception for robot systems, particularly in the 3D scene understanding that can bridge perception to planning and control. His current focus is to explore efficient approaches for dense 3D motion (scene flow) from videos, which is the core of the thesis topic.