DA-Net: Density-Adaptive Downsampling Network for Point Cloud Classification via End-to-End Learning

Since a point cloud might contain large quantities of points in practical scenarios, it is desirable to perform downsampling before point cloud analysis. Classic downsampling strategies, such as farthest point sampling (FPS), tend to sample points regardless of subsequent tasks, which inevitably hampers the effectiveness of sampling results. Moreover, raw point clouds obtained from 3D sensors are usually noisy and unevenly distributed. As a consequence, the downsampling process might be affected seriously. In this paper, we propose a density-adaptive downsampling network (DA-Net) for point cloud classification task. The entire process of DA-Net is conducted in an end-to-end way, ensuring that the sampling operation is differentiable and optimal for downstream applications. One key process of DA-Net is density-adaptive K nearest neighboring (DAKNN), which is proposed to alleviate the adverse effects brought by density variation of point clouds. Another key process is the local adjustment (LA) of initial sampled points, which can further equip our model with noise immunity. Consequently, our network can enhance the effectiveness of sampled points and further improve the task performance. In point cloud classification task, our approach achieves significantly better results over existing downsampling alternatives. Extensive experiments are presented to demonstrate the advantages of our DA-Net.