Abstract
As the light detection and ranging (LiDAR) sensor can precisely sense the distance between sensors and the environment. It has placed great expectations on 3D scene understanding. However, point clouds generated by LiDAR are always sparsely distributed in the 3D space with unstructured storage, which makes it hard to represent learning. This paper proposes a surface focusing transformer one-stage 3D object detection method. In this method, this paper designs a surface focusing transformer as the 3D encoder to efficiently learn representative information from point clouds. During the downsampling process of the point cloud, an instance-aware downsampling strategy will be adopted to allow our method to pay more attention to the foreground areas. Extensive experiments demonstrate that our method can suppress other one-stage 3D detection methods with a clear margin.