Boosting 3D Object Detection with Density-Aware Semantics-Augmented Set Abstraction

In recent years, point cloud-based 3D object detection has seen tremendous success. Previous point-based methods use Set Abstraction (SA) to sample the key points and abstract their features, which did not fully take density variation into consideration in point sampling and feature extraction. The SA module can be split into three parts: point sampling, grouping and feature extraction. Previous sampling methods focus more on distances among points in Euclidean space or feature space, ignoring the point density, thus making it more likely to sample points in Ground Truth (GT) containing dense points. Furthermore, the feature extraction module takes the relative coordinates and point features as input, while raw point coordinates can represent more informative attributes, i.e., point density and direction angle. So, this paper proposes Density-aware Semantics-Augmented Set Abstraction (DSASA) for solving the above two issues, which takes a deep look at the point density in the sampling process and enhances point features using onefold raw point coordinates. We conduct the experiments on the KITTI dataset and verify the superiority of DSASA.