Multi Point-Voxel Convolution (MPVConv) for deep learning on point clouds

The existing 3D deep learning methods adopt either individual point-based features or local-neighboring voxel-based features, and demonstrate great potential for processing data. However, the models are inefficient due to unordered nature of point clouds suffer from large information loss. Motivated by success recent point-voxel representation, such as PVCNN DRINet, we propose a new convolutional neural network, called Multi Point-Voxel Convolution (MPVConv), on clouds. Integrating both advantages voxel methods, MPVConv can effectively increase neighboring collection between also promote independence among features. Extensive experiments benchmark datasets ShapeNet Part, S3DIS KITTI various tasks show that improves accuracy backbone (PointNet) up 36%, achieves higher than model with 34× speedups. In addition, outperforms state-of-the-art 8× Also, our only needs 65% GPU memory required latest point-voxel-based (DRINet). source code method is attached in https://github.com/NWUzhouwei/MPVConv.