CA-Net: contextually aware point cloud downsampling network using LSTM

Point cloud data often has a large volume, and it is often sampled to reduce computational load and storage. Common sampling methods like Random Sampling (RS) and Farthest Point Sampling (FPS) can rapidly sample point cloud. However, these sampled points are not related to subsequent tasks. Current task-driven approaches generate sampled point cloud in batches regardless of contextual information. Moreover, they ignore the distance between sampled points, which results in dense distributions. To address these issues, we propose a point cloud downsampling network based on Long Short-Term Memory (LSTM), named CA-Net. It uses three-layer DGCNN and mixed pooling to achieve feature vectors, a LSTM network to catch contextual information, and two sampling loss functions to make the sampled points uniformly distributed. Experimental results on reconstruction and classification tasks demonstrate that the proposed downsampling network significantly outperforms existing methods.