Learning-Based Sampling Method for Point Cloud Segmentation

Light detection and ranging (LiDAR) has become one of the most important sensors in 3-D perception. With the advancement of sensor technology, the point cloud data generated by LiDAR have also become increasingly large. There are difficulties in processing them directly, such as hardware limitations, computational costs, storage constraints, algorithmic considerations, and other factors. One of the most common solutions is to sample the point clouds. Learning-based downsampling methods have been proven to be effective in point cloud classification, registration, and reconstruction. However, the integration of downsampling techniques with segmentation tasks remains inadequately investigated. This is mainly because, for segmentation tasks, to achieve higher segmentation accuracy, the sampled points need more detailed information and complete structural information. This will greatly increase the difficulty of sampling. This article proposes a learning-based sampling method for point cloud segmentation task. Our research analyzes the spatial relationships within point clouds using a simplification network to generate sampled points. A bidirectional chamfer distance (CD) is used to ensure that the original and sampled points have similar structural characteristics. The experimental results demonstrate that our network, SampleSegNet, outperforms alternative sampling methods.