Feature Distance Representer: Exploring the Connection Between Point Cloud Feature Levels From the Perspective of Distance

Kaijun You, Hou zhenjie, Jiuzhen Liang, Hao Liu

Abstract: With the advent of the Industry 4.0 era, intelligent driving, medical impact, industrial automation, and other fields have developed rapidly, point cloud acquisition technology and point cloud identification network have been widely used, and point cloud data in the field of sensor intelligent identification are also becoming more and more important. Point cloud data are generated by 3-D sensors. Previous processing of this type of data has mainly relied on the use of convolution, graphs, or attention mechanisms to explore complex local geometric features. However, in the process of exploration of these methods, few people have explored the internal relationship between different levels of features, and revealing the relationship can make us more clearly understand the mutual relationship between features. Based on the above considerations, we characterize the connection between different levels of features by distance, which is different from the traditional study of spatiotemporal dimensional features, and we enhance the discrepancy of different object classifier features from the perspective of the connection between different levels of features. Usually, the measure of such variability is the cosine similarity. The cosine similarity distinguishes the difference more from the direction, it is impossible to measure the difference of the value in each dimension, while the kernel tensor can represent the connection between each dimensional vector component. Therefore, we propose a feature distance representer (FDR) under Hilbert space. We embed point features into the Hilbert space and compute the distance between classifier features and core tensor features of point cloud in the Hilbert space to characterize the interrelationships between hierarchical features, which leads to substantial improvements in the performance of state-of-the-art network models for point cloud classification. The average classification accuracy reached up to 92.23% on PointNet++, up to 90.79% on dynamic graph convolutional neural network (DGCNN) and up to 94.00% on PointMLP which is the latest residual network. Our code is released at: https://github.com/yyykj/FDR-master.