Learnable Upsampling-Based Point Cloud Semantic Segmentation

The point cloud semantic segmentation network based on point-wise multi-layer perceptron (MLP) has been widely applied with its end-to-end advantages. Normally, such networks use the traditional upsampling algorithm to recover the details of point clouds in the decoding stage. However, the point cloud has rich 3D geometric information. The traditional interpolation algorithm does not consider the geometric correlation in the process of recovering the details of the point cloud, resulting in the inaccurate output point features. To this end, a learnable upsampling algorithm is proposed in this paper. This upsampling algorithm is implemented by utilizing moving least squares (MLS) and radial basis function (RBF), which can fully exploit the local geometric features of point clouds and accurately restore the details of scenarios. The validity of the proposed upsampling operator is verified on the Semantic3D dataset. Experimental results show that the proposed upsampling algorithm is superior to the widely applied traditional interpolation algorithms when used for point cloud semantic segmentation.