Recovery of Sharp Feature Based on Neighborhood Fused PointNet

Feature points are important elements of the model, and how to use uniform parameters to extract feature points from multi-scale feature point clouds is a challenging problem. Therefore, this paper proposes a feature point extraction method based on a data-driven point cloud.This method can extract the feature points of multi-scale feature point cloud by unifying the model parameters. Specifically, we fuse the center feature and the neighborhood multi-channel feature, and map the category of the corresponding center point through the neural network. In addition, based on the reliable model feature points, we propose a novel joint loss function based on the expected position and normal vector constraints, which allows the smooth features to recover the sharp state. Experiments show that our algorithm can accurately extract feature points from the model point cloud with strong noise and multi-scale features, and realize the reconstruction of sharp features on feature points with good robustness.