A novel plane based image registration pipeline with CNN scene parsing

Plane is one of the most important element for indoor man-made structural rooms, such as broadcasting studios, open lecture room and empty offices. The existing visual mapping algorithms cannot effectively detect and describe the visual features on an empty large or medium size plane. This article aims to introduce a novel frame-to-frame registration pipeline based on one medium-size plane on man-made object instead of multiple planes or small plane patches. By introducing a structural description of reference planar area with its contour data and CNN segmentation information, the proposed approach is able track the pose of camera with high accuracy and robustness in comparison with existing feature-based tracking or dense geometric tracking approaches.