Event Camera-based Motion Segmentation via Depth Estimation and 3D Motion Compensation

Motion segmentation is the basic step in many computer vision works. Its purpose is to separate the moving foreground from the background in a video. In this paper, we propose a motion segmentation algorithm by fusing depth estimation and a three-dimensional (3D) motion compensation approach. The pseudo-depth of per-event in each frame of pixel is obtained by depth estimation, linear interpolation, and coordinate transformation. The Volume of Warped Events (VWE) is conducted by warping the events and iterating the parameters, and a 3D motion segmentation algorithm is presented to maximize the contrast of VWE. In experiments, we establish our dataset by decomposing a video sequence into consecutive frames, which can be utilized in ESIM to obtain the dataset. The experimental results validate the feasibility of the proposed algorithm, and show the maximal contrast of VWE in our dataset.