Wave-iPCRNet: Toward point cloud registration in electronics manufacturing by a quantum-inspired iterative point cloud registration network

Point cloud registration is one of the important 3D vision tasks. It is a fundamental task of the downstream tasks such as 3D tracking, autonomous driving, 3D reconstruction, and pose estimation. The iterative point cloud registration network (iPCRNet) model is developed by the team of Carnegie Mellon University et al., using point cloud data directly to perform the point cloud registration task. On the other hand, the defect detection in the electronic manufacturing has the requirements of minimal inference time and training cost. Despite the transformer module has achieved good performance on this task, its computation cost increase rapidly while the input data points increased than other modules. Hence, considering these requirements the multi-layer perceptron (MLP) module is usually used. However, the simple MLP module performance on the network design may have some improvements can be done. This work proposed a Wave-iPCRNet model using the quantum-inspired Wave-MLP module achieving the state-of-the-art performance on numerous 2D tasks. On the ModelNet40 benchmark dataset, the Wave-iPCRNet improves the test loss of iPCRNet from 0.038 to 0.031, improving the rotation error of iPCRNet from 15.153 to 14.104, and improving the translation error of iPCRNet from 0.007 to 0.006.