摘要:近年来,视觉Transfommer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Tansfomer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人表示,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制模型PoiniConvMixer,用ConyMixer架构取代了 Poin-BERT中的标准Transformer。PointConvMlixer在 ModelNet40数据集上的整体分类准确率达到92.3%,在ShapeNetPars数据集上进行点云部分分制时mI0Ul和mOUC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 PPFConvhlixer,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.PPFConwMixer的总体分类准确率达到了93.8%。

Abstraet: In recent years, Vision Transformers (ViTs) show potential supericrity on 3D computer vision tasksincluding point cloud classification, but the provenance of their eflectiveness remains ambiguous. lt is highlyessential to investigate whether their perlormance in vision tasks is entirely due to the superiority of the structureitself, or at least partially benefits from the use of local patches as input representations. Motivated by this, in thispaper PointConvMixer was proposed, a simple but still effective point cloud classification and segmentationmodel, replacing the standard Transformer in Point-BERT with the ConvMixer architecture. The overallclassilication accuracy of PointConvMixer on the vodelNet40 dataset reaches 92,3%, and the mlOUl and mlOUcfor point cloud segmentation on the ShapeNet Parts dataset are 85.4% and 83.9% respectively, both of whichoutperlorm the compared Transformer-Based networks. In addition, PpFConvixer was further introduced. whichaugmented PointConvlMixer with an eficient local feature deseriptor Point Pair Feature (PpF) to optimize thepoint cloud classilication perlormance. Our method has shown promising results for point cloud analysis despitcits simplicity.The overall classilication accuracy ofPpFConvMixer achieves 93.8% at a query radius of0.25m.