WireframePose: Monocular 6-D Pose Estimation of Metal Parts Based on Wireframe Extraction and Matching

摘要

The 6-D pose estimation of metal parts is an essential task of vision-based measurement for precise robot grasping and assembly in unstructured scenes. However, the textureless nature of metal parts presents challenges in extracting effective visual features for accurate pose estimation. To tackle this problem, considering that wireframe, which refers to straight lines and their endpoints, is independent of texture, we propose a novel two-stage pose estimation method based on wireframe extraction and matching called WireframePose. The whole framework only requires lines and endpoints for pose estimation, which overcomes the reliance on texture information. In the first stage, a multitask network is proposed to predict the wireframe of metal part with a monocular image. In the second stage, a robust descriptor constructed by the relative angles of the adjacent lines is proposed to match the predicted wireframe with the templates. The matched endpoints establish the 2-D–3-D correspondences for estimating pose. Furthermore, to minimize the domain gap during matching, a training data and wireframe templates generation method based on the projection of 3-D model wireframe is proposed. Experimental results demonstrate that the proposed method achieves competitive results with a real-time running speed for the pose estimation of metal parts.