Asfcnet : Adaptive Sampling and Feature Capture for Point Cloud Processing

In recent years, semantic segmentation and classification of 3D scenes based on point cloud data have become a hot research topic. A large number of deep learning methods have been developed for point cloud. However, there are at least two problems in point cloud deep learning neural networks. (1) Existing works widely use max/mean pooling to integrate the neighbor features, which will cause the majority of the information to be lost. (2) Point cloud data generally contains many noises, and the widely adopted farthest distance sampling is easily interfered by these noises. This paper proposes a feature capture module, which is to capture local features and global features. The local feature capture module is inspired by Taylor's formula. It learns the structure of neighbor points within a group in a low-dimensional space. The global feature capture module learns the global information by fusing the features between groups. In addition, we reduce the impact of noise on the network from two aspects. Firstly, an adaptive sampling module is proposed to adjust the position of sampling points. Secondly, an adaptive interpolation module is developed to learn the interpolation weights of different points in the point feature propagation. The experiments are carried out on widely used point cloud datasets S3DIS, Shapenet, and Modelnet40. AsfcNet network can handle multi-tasks and achieve excellent experimental results compared with many single task networks