【摘要】 基于体素的三维目标检测模型在速度和检测精度上要超过基于图的模型，但是在体素化过程中使用平均池化操作会导致细节信息的丢失，从而在一定程度上降低模型的性能。论文使用图网络在体素化过程中显式构造拓扑结构捕获局部点云细节信息来解决体素化中的信息丢失问题，并通过裁剪体素骨干网络达到速度和检测精度的平衡。提出的方法在公开的三维目标检测数据库KITTI上进行了汽车类别目标的检测实验，取得了84.85%的均值精度（AP）检测结果，超过了一些先进的三维目标检测模型。更多还原

【Abstract】 The voxel-based 3D object detection model outperforms the graph-based model in terms of speed and detection accuracy,but the use of average pooling operation during voxelization leads to loss of detail information,which degrades the performance of the model to some extent. The proposed method uses a graph network to explicitly construct the topology to capture local point cloud detail information during voxelization to solve the information loss in voxelization,and achieves a balance of speed and detection accuracy by cropping the voxel backbone network. The proposed method is experimented on the KITTI,a publicly available 3D object detection database,for the detection of car class objects,and achieves 84.85% average precision(AP)detection results,which exceedes some advanced 3D object detection models.更多还原