Rethinking Normals: Direction Guided Point Cloud Recognition

A point cloud is a geometric data structure that is a set of coordinates p(x, y, z) of points, with additional normal n(n x , n y , n z ) data. The current popular point-based method mainly deals with coordinates (x, y, z), while normals (n x , n y , n z ) are used as additional data channels. Since normals represent direction information and coordinates represent position information, they are different types of data. Directly adding normal data to coordinate information will cause the direction information to be weakened, resulting in a low accuracy improvement effect of the model, generally only around 0.2%∼0.8%. To fully utilize the normal information of point clouds, we have designed a Direction-Based FPS(farthest point sampling) algorithm, which can reduce the detail loss caused by point cloud downsampling; At the same time, we propose a method of Direction-View features from normal data and fuses them with Point-View features to improve the recognition ability of the base model. Experiments show that our designed method can effectively utilize the direction information features of normals, improving the classification accuracy of the base model on the ModelNet40 dataset by more than 2%.