基于ConvMixer架构的高效点云分类方法  

An efficient point cloud classification method based on ConvMixer architecture

作　　者：王淳 赵艳明[1] 冯燕 WANG Chun;ZHAO Yanming;FENG Yan(School of Information and Communication Engineering,Communication University of China,Beijing 100024,China)

机构地区：[1]中国传媒大学信息与通信工程学院,北京100024

出　　处：《中国传媒大学学报（自然科学版）》2024年第1期56-64,共9页Journal of Communication University of China：Science and Technology

基　　金：国家重点研发计划(2018YFB1404103)。

摘　　要：近年来,视觉Transformer模型在点云分类等三维计算机视觉任务中显现出潜在的优越性,但其有效性来源仍然模糊不清。研究它们在视觉任务中的性能是完全归功于Transformer结构本身的优越性,还是至少部分得益于使用局部块作为输入表示,是非常必要的。受此启发,本文提出了一种简单但仍然有效的点云分类和分割模型PointConvMixer,用ConvMixer架构取代了Point-BERT中的标准Transformer。PointConvMixer在ModelNet40数据集上的整体分类准确率达到92.3%,在ShapeNet Parts数据集上进行点云部分分割时mIOUI和mIOUC分别为85.4%和83.9%,均优于基于Transformer的对比模型。此外,本文还进一步提出PPFConvMixer,其利用高效的局部特征描述符PPF增强了PointConvMixer,从而优化了点云分类性能。在查询半径为0.25m时,PPFConvMixer的总体分类准确率达到了93.8%。In recent years,Vision Transformers(ViTs)show potential superiority on 3D computer vision tasks,including point cloud classification,but the provenance of their effectiveness remains ambiguous.It is highly essential to investigate whether their performance in vision tasks is entirely due to the superiority of the structure itself,or at least partially benefits from the use of local patches as input representations.Motivated by this,in this paper PointConvMixer was proposed,a simple but still effective point cloud classification and segmentation model,replacing the standard Transformer in Point-BERT with the ConvMixer architecture.The overall classification accuracy of PointConvMixer on the ModelNet40 dataset reaches 92.3%,and the mIOUI and mIOUC for point cloud segmentation on the ShapeNet Parts dataset are 85.4%and 83.9%respectively,both of which outperform the compared Transformer-Based networks.In addition,PPFConvMixer was further introduced,which augmented PointConvMixer with an efficient local feature descriptor Point Pair Feature(PPF)to optimize the point cloud classification performance.Our method has shown promising results for point cloud analysis despite its simplicity.The overall classification accuracy of PPFConvMixer achieves 93.8%at a query radius of 0.25m.

关 键 词：三维点云分类 深度学习 ConvMixer Point Pair Feature 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]