GDS: Global Description Guided Down-Sampling for 3D Point Cloud Classification

In deep neural networks for 3D point clouds, the down- sampling operation is a key module for effectively improving the computational efficiency as well as boosting robustness to variation of input points. Previous works mainly utilize furthest point sampling to down-sample points in accordance with the spatial distance between points, which is an offline and time-consuming operation. In this paper, we propose a Global Description guided down-Sampling method (GDS) to learn to sample points from the input point set in accordance with their features. Specifically, through retaining points features with high affinity to the global shape description, our GDS module preserve significant points features and their coordinates on the fly. We also equip our GDS with a locality feature aggregation module to form Global Description guided Pooling operation (GDP) for 3D point networks. Experimental results on two publicly available datasets, ModelNet and ScanObjectNN, show that introducing the proposed GDS and GDP into 3D object classification networks can effectively reduce about 45% of the forward propagation time while achieving higher accuracy.