Accurate pose estimation of a hand-held RGBD camera based on sub-volume matching for 3D modeling

The smoothness of the result of full body 3D reconstruction, also known as 360^◦ reconstruction using a single hand-held sensor depends on the accuracy of the pose estimation. In this paper, we present a new idea for accurate pose estimation of such a single hand-held RGBD sensor based on subvolumetric reconstruction. In our method, we first estimate initial pose of both RGB and depth sensors through 3D coarse registration. Thereafter, in the precision matching step, we select only the keyframes for matching and estimate relative pose between them based on registration refinement. If there is a large pose estimation error between the keyframes, a subvolume is constructed using data of adjacent frames of each keyframe, and refine the final relative pose between keyframes using subvolume estimations. A series of 3D reconstruction experiments are preformed to evaluate the accuracy of the estimated pose.