PVS-CNN:POINT-VOXEL CNN OPTIMIZED BY SUBMANIFOLD SPARSE CONVOLUTION
In order to solve the problem of low efficiency and large GPU memory usage in segmentation and detection of classic 3D convolutional networks in scenes with larger models,this paper proposes the PVS-CNN network framework,which achieves 3D convolution with high efficiency and low GPU occupancy by updating the Hash table and feature sparse matrix,and uses submanifold sparse convolution to improve PV-Conv.The PVS-CNN was evaluated on ShapeNet and S3DIS dataset.The experimental results show that the proposed PVS-CNN is 3.6 times faster than PVCNN and the GPU memory usage is only 0.55 times that of PVCNN.Compared with F-PVCNN in object detection,the proposed PVS-CNN is better than F-PVCNN in terms of time efficiency and detection accuracy.
Three-dimensional point cloudEfficiencyGPU memory usageSegmentationObject detectionSubmanifold sparse convolution