Layered RGBD Scene Flow Estimation with Global Non-rigid Local Rigid Assumption

Abstract

RGBD scene flow has attracted increasing attention in the computer vision community with the popularity of depth sensor. To accurately estimate three-dimensional motion of object, a layered scene flow estimation with global non-rigid, local rigid motion assumption is presented in this paper. Firstly, depth image is inpainted based on RGB image due to original depth image contains noises. Secondly, depth image is layered according to K-means clustering algorithm, which can quickly and simply layer the depth image. Thirdly, scene flow is estimated based on the assumption we proposed. Finally, experiments are implemented on RGBD tracking dataset and deformable 3D reconstruction dataset, and the analysis of quantitative indicators, RMS (Root Mean Square error) and AAE (Average Angular Error). The results show that the proposed method can distinguish moving regions from the static background better, and more accurately estimate the motion information of the scene by comparing with the global rigid, local non-rigid assumption.