【摘要】 针对类级别的物体姿态估计问题，提出一种仅将深度相机扫描的点云作为输入，在仅知道目标物体点云类别的情况下，准确估计目标物体三维位姿的方法。该方法不需要依赖大量的带标签的人工标注数据集，仅使用虚拟仿真技术模拟生产的数据，即可在真实数据集上取得较高的精度。该方法首先对输入点云进行背景噪声过滤，之后通过中心预测模块对点云做标准归一化，再使用基于对应类别模板点云变形的方法预测其标准坐标系坐标，最后通过最小二乘法获得目标物体的三维位姿。实验结果表明，该方法在真实数据上具有更好的泛化性能和更高的精度。更多还原

【Abstract】 Aiming at the problem of category-level object pose estimation, a method was proposed to accurately estimate the pose of the target object by only taking the point cloud scanned by the depth camera as the input, with knowing the category of input point cloud only. The method did not reply on a huge amount of labeled dataset, but used virtual data produced by simulation instead, which achieved better accuracy on real-world dataset. This method first filtered the background noise of the input point cloud. Then standardized the point cloud through the well-designed center prediction module. After that, the normalized object coordinate space would be estimated through a shape template deformation module. Finally, the pose would be obtained from least squares. Experiments on real-world dataset demonstrates that the method achieve higher accuracy and better generalization ability.更多还原