Voxel Transformer with Shifted Windows for 3D Object Detection

Recent three-dimensional object detection methods are typically classified into point-based and voxel-based categories based on the processing method of raw point clouds. Voxel-based methods, which convert the point clouds to voxels to reduce computational load, often suffer from the geometric information loss and limited detection accuracy. In this paper, we propose a novel single-stage and voxel-based 3D object detection algorithm (VWTr) using Voxel Feature Encoder to extract features and Transformer Backbone with shifted windows to enhance the capability of feature extraction, which achieves a balance between accuracy and speed. The Transformer Backbone with shifted windows can help the network efficiently concentrate on global information and make up for the geometric information loss arose from the voxelization operation of the voxel feature encoder. To this end, we design a feature aggregation operation to enhance the network's representation capability. Relevant experiments on KITTI have demonstrated that our method has respectively reached 84.11%, 75.18%, 69.53% AP on three difficulty levels of KITTI car test set, which is superior to PointPillars algorithm,