PFSC:Pyramid R-CNN for Point-Voxels with Focal Sparse Convolutional Networks for 3D Object Detection

In the point cloud 3D object detection task, the point cloud data is unevenly distributed in 3D space. Therefore, whether it is regular sparse convolution or submanifold sparse convolution, there is a significant difference in the detection performance when dealing with objects at different positions using sparse convolutional networks. Therefore, this paper proposes an improvement using focal sparse convolution to break the limitations of traditional convolution methods and enhance the representation ability of sparse convolution. Focal sparse convolution is utilized to replace the currently better-performing two-stage object detection network based on the combination of points and voxels, enhancing the end-to-end training effectiveness and achieving better detection results. And an optional detachable auxiliary network is added, which, during the training stage, combines point features to guide the backbone network in learning the structural information of the 3D point cloud, enhancing the localization accuracy of the backbone network. Extensive experiments have been conducted on the KITTI dataset to validate the proposed method, yielding promising detection results and demonstrating the effectiveness of the approach.