A Single-Stage 3D Object Detector with Multi-scale Input and Output

3D object detection plays an important role for autonomous vehicles and a number of models have been proposed to handle this task. Among them, PointPillars has an outstanding advantage on fast encoding point clouds. However, the 3D structural information are always weaken by the voxelization. Furthermore, the single-scale output seriously limits the detection accuracy on the objects of different sizes. Motivated by that, this paper proposes a single-stage object detection network with multi-scale input and output. First, the grids of different scales are introduced to voxelize the point cloud and generate feature maps of different resolutions. Then, the maps are processed by an efficient fusion network. Finally, a multi-scale detection layer is designed to obtain 3D bounding boxes and categories. Comparative and ablative experiments demonstrate that our scheme can effectively improve the performance of 3D object detection.