Pose Refinement of Occluded 3D Objects Based on Visible Surface Extraction

Abstract

In this paper, we propose a pose refinement method based on the visible surface extraction of 3D object. Given a rough estimation of object pose, the algorithm of iterative closet point (ICP) is often used to refine the pose by aligning the object model with test scene. To avoid the interference of invisible points on the ICP process, we only use the visible surface for pose refinement. It is especially necessary when occlusion occurs in the scene. Combining the technologies of image rendering and depth consistency verification, the visible surface can be effectively extracted. During the process of pose refinement, hypothesis verification methods are also used to eliminate unreasonable hypothetical poses as early as possible. The proposed method is evaluated on the public Tejani dataset. The experimental results show that our method improved the average F1-score by 0.2062, which proves that our method can obtain pose estimation results of high accuracy, even in the occluded scene.