Quantifying point cloud realism through adversarially learned latent\n representations

Abstract

Judging the quality of samples synthesized by generative models can be\ntedious and time consuming, especially for complex data structures, such as\npoint clouds. This paper presents a novel approach to quantify the realism of\nlocal regions in LiDAR point clouds. Relevant features are learned from\nreal-world and synthetic point clouds by training on a proxy classification\ntask. Inspired by fair networks, we use an adversarial technique to discourage\nthe encoding of dataset-specific information. The resulting metric can assign a\nquality score to samples without requiring any task specific annotations.\n In a series of experiments, we confirm the soundness of our metric by\napplying it in controllable task setups and on unseen data. Additional\nexperiments show reliable interpolation capabilities of the metric between data\nwith varying degree of realism. As one important application, we demonstrate\nhow the local realism score can be used for anomaly detection in point clouds.\n