Abstract: Aiming at the problems that traditional vision algorithms cannot accurately measure the 3D center of mass and size of an object, and cannot detect the objects that have not appeared in the dataset, we propose a category-level object center of mass and position estimation method based on point cloud. Firstly, a 2D segmentation algorithm is used to segment the region to be detected, and BlendMask is used to improve the real-time performance; then, a point cloud is generated from the segmented region and combined with the a priori shapes to reconstruct the object model in a neural network; finally, a deformation field and its corresponding matrix are outputted, and the position of the object is estimated by the Umeyama algorithm. The experimental results show that the proposed method achieves high accuracy, outperforms the center-of-envelope-box algorithm in center-of-mass estimation, and achieves a real-time rate of 13 FPS. Simulation experiments on a GPU NVIDIA 2060 demonstrate that the center-of-mass of the proposed method outperforms the comparative methods, and that the method is both accurate and reliable.