ScaleLK: Registration of Point Clouds with Different Scales Using Deep Learning Methods

Abstract

3D point clouds are widely used in numerous research and applications, such as autonomous driving, industrial robots, and augmented reality, to represent the spatial structure of objects. The 3D point cloud registration aims to transform the source point cloud into the same coordinate system with the template point cloud, which is of great significance for the 3D reconstruction of the real world objects. ICP [1] is one of the most classic point cloud registration algorithms but it still has problems with efficiency and initialization. With the help of deep learning, PointNetLK [7] becomes a state-of-the-art point cloud registration method. Although PointNetLK is efficient and robust to some extent, it is not able to register point clouds with different scales. In this paper, we propose ScaleLK, an approach for registration of point clouds with different scales using deep learning methods. We have trained a feature extractor which supports scale feature and used this feature for registration of point clouds with different scales. We describe the architecture and compare its performance with other methods.