Efficient Object Localization for Unseen Object 6D Pose Estimation

Abstract

Object localization is utilized as the first step in standard 6D object pose estimation methods to obtain the position information of the objects. However, these object localization methods cannot be directly applied to unseen objects, which is the focus of recent research on 6D object pose estimation. In this paper, an accurate and efficient localization method for unseen object is proposed, based on a template matching strategy. The Hybrid Channel-Spatial Attention Model (HCSAM) is designed to focus on the target object by enhancing the contextual differences between the target object and background. Additionally, The Multi-Scale Integration Transformer (MSIT) module is designed to eliminate noise interference and enhance semantic information in low-dimensional features by integrating multidimensional information. Our method outperforms existing approaches on the complicated occluded dataset LINEMOD, as well as on the challenging generalized pose estimation dataset GenMOP.