OA-Pose: Occlusion-aware monocular 6-DoF object pose estimation under geometry alignment for robot manipulation

Abstract

Object pose estimation is the fundamental technology of robot manipulation systems. Recently, various learning-based monocular pose estimation methods have achieved outstanding performance by establishing sparse/dense 2D-3D correspondences. However, in cluttered environments, occlusion has been a challenging problem for pose estimation due to limited information provided by visible parts. In this work, we propose an efficient occlusion-aware monocular pose estimation method, called OA-Pose, to learn geometric feature information of occluded objects from cluttered scenes. Our framework takes RGB images as input and generates 2D-3D correspondences of visible and invisible parts based on the proposed Occlusion-aware Geometry Alignment Module. Extensive experiments show that our method is superior and competitive with state-of-the-art on multiple public datasets. We also conduct grasping experiments with different degrees of object occlusion, demonstrating the usability of our algorithm to deploy on robots in unstructured environments.