Evolutionary Neural Architecture Search Framework With Masked Encoding Mechanism for Point Cloud Registration

Deep learning-based point cloud registration (PCR) has received increasing attention in the field of computer vision. However, the lack of expertise in deep neural network architecture design remains a large obstacle for end-users with registration needs. While neural architecture search (NAS) indeed addresses the architecture automation design problem and has undergone substantial development, the majority of existing NAS approaches are tailored for image processing. Because of significant differences in data characteristics and network structures, directly transferring the existing NAS approaches for PCR is infeasible. Hence, this work proposes an evolutionary neural architecture search framework for PCR, which can fit the different types of PCR networks. The framework, equipped with the designed masked encoding mechanism, can seamlessly integrate with the majority of evolutionary computation-based optimization algorithms, thereby reducing the additional cost of operator design. Based on the above framework, the users can freely select the advanced network components and optimization algorithms depending on their needs. Experimental results on synthetic and real-world datasets, ModelNet40 and KITTI, demonstrate the superiority of the proposed framework in terms of registration precision and architecture complexity.