Centroid and pose estimation of occluded objects based on deep learning

Aiming at the problems that traditional visual algorithms can't accurately measure the 3D centroid and size of objects and can't detect objects that haven't appeared in the data set, this paper proposes a point cloud-based centroid and pose estimation method for class-level objects.  At the same time, the method mentioned in this paper can greatly reduce the production cost of data set.  The unseen object instances of the same category are recovered from a single RGBD image, i.e. the instances that do not exist in the data set, so as to estimate their centroid and pose.  The object category is first determined by a two-dimensional algorithm and the detection area is divided. The observed point cloud is obtained by combining with the depth map. The RGB image, observed point cloud and prior shape are input into the network to output a deformation field and corresponding matrix.  The prior shapes were deformed and then normalized to NOCS coordinates, and registration was carried out by Umeyama algorithm.  Thus, the class and pose of the object are determined, and the centroid of the object is calculated. Meanwhile, BlendMask is used for two-dimensional segmentation to realize real-time pose estimation.  Simulation results show the accuracy and reliability of the proposed method, and the point cloud method is superior to other methods in calculating the center of mass.