A study of point clouds based on geometric disentanglement and feature fusion

The point cloud can be de-entangled into sharply varying frame components and slowly varying plane components, which complement each other to construct complete point cloud geometric information. Geometric de-entanglement attention network (GDAnet) dynamically de-entangles the point cloud into the contour and plane portions of a 3D object, which are represented by sharply varying and slowly varying components, respectively, and supplements the localized information by capturing and refining the overall and complementary 3D geometric semantic meanings. In this paper, the geometric disentanglement part and the feature fusion part of GDAnet are improved, and the metric formula is improved in the disentanglement part to make the disentanglement better, while the attention calculation and fusion and splicing operation of the components reduces some of the redundancy information, fully explores the relationship between the points, and incorporates the remote context into the local information, which makes the improved network faster and more accurate in the ModelNet40 and SharpNet datasets are significantly improved.