Point Cloud Distance Metrics for Evaluation of Deep Point Networks

Point cloud processing plays a crucial role in 3D representation and modeling. Metrics such as Chamfer Distance (CD) and Earth Mover's Distance (EMD) are extensively used in point cloud processing, for instance, in evaluating the point clouds reconstructed by point Autoencoders (AEs). It has been reported that the choice of metric significantly influences the performance of point cloud processing. We have discovered that the degree of one-to-one correspondence embedded in the distance metrics governs the error tendency, such shape deformations as point dispersion and aggregation, in the point clouds reconstructed by point AEs. This discovery has prompted us to propose the Modified Chamfer Distance (MCD) metric, which allows us to control the level of one-to-one correspondence in-between CD and EMD, enabling us to achieve an optimal error tendency for a given problem. Experimental results validate the effectiveness of the proposed MCD.