IndexSample: A Learnable Sampling Network in Point Cloud Classification

Recently, growing attention has been paid in the process of point cloud with deep learning method. It is natural to process these data hierarchical as 2D convolutional neural network(CNN) suggests, which needs to downsample the point clouds. Classic sampling approaches, such as farthest point sampling (FPS), do not consider the downstream task. Recent works showed that learning a task-specific sampling method can improve results for different tasks. We introduce a novel differentiable sampling method called IndexSample and its corresponding loss constrain. With sampled points, we add a confidence layer to further balance the sampled points' weight. During the sampling task, we have to make sure the sampling covers the most remarkable features and keep the result consistent between different point clouds, which is challenging. Experiments show that our sampling method achieves on par or better performance than classic sampling methods with much faster speed.