A Visualization-aware Loss for Point Cloud Comparison

Autonomous driving is a worthwhile and popular field in both the industry and the academia community, with point clouds commonly utilized. Recently, numerous methods have been proposed to model point clouds for subsequent applications. However, little attention has been paid into the fundamental research of point cloud comparison. Objectively speaking, point cloud comparison plays an irreplaceable role in point cloud modeling. The main significance is that it is the general guidance and assessment for any method. Recently, modeling related to the human interaction and simulation is the prevailing research tendency, highlighting the significance of visual effects. However, current comparison metrics including Chamfer Distance and Earth Movig Distance are lacking in the consideration of the visualization effects. Therefore, we first give an intuitive visual comparison and analysis of those two distances. On this basic, we propose our visualization-aware loss with a structural matching scheme. Our loss is composed with the Max-Resample Chamfer Distance (MCD) and the Matching Consistency Loss (MCL). Specifically, MCD is a new metrics, which removes redundant point pairs to be more sensitive to the maintenance of the global visualization effect. As a training auxiliary, we originally propose a multi-scale MCL, which can apparently reduce the visible artifacts. Quantitative and qualitative experiments have demonstrated the exactness and the effectiveness of our distance metrics and auxiliary loss on several point cloud networks and datasets.