3D Semantic Segmentation Based on Improved Residual Block Point-Voxel Feature Fusion Network

This paper proposes a 3D semantic segmentation network for point cloud voxel based on improved residual blocks. In traditional methods, point clouds are usually projected into two-dimensional space and processed by two-dimensional convolution. This method inevitably loses the connection between the points. However, the disorder of point cloud and the randomness of each position are not convenient for direct convolution, and voxelization cannot avoid the loss of information. Therefore, we designed the fusion method of point cloud and voxel feature for semantic segmentation, and improved the residual block, enhanced the feature extraction, and increased the training speed, and got more accurate results. Experimental results on the SemanticKITTI large-scale dataset demonstrate the validity of our proposed 3D semantic segmentation network.