Point cloud analysis is challenging due to the irregularity of the point cloud data structure. Existing works typically employ the ad-hoc sampling-grouping operation of PointNet++, followed by sophisticated local and/or global feature extractors for leveraging the 3D geometry of the point cloud. Unfortunately, those intricate hand-crafted model designs have led to poor inference latency and performance saturation in the last few years. In this paper, we point out that the classical sampling-grouping operations on the irregular point cloud cause learning difficulty for the subsequent MLP layers. To reduce the irregularity of the point cloud, we introduce a DualNorm module after the sampling-grouping operation. The DualNorm module consists of Point Normalization, which normalizes the grouped points to the sampled points, and Reverse Point Normalization, which normalizes the sampled points to the grouped points. The proposed PointNorm utilizes local mean and global standard deviation to benefit from both local and global features while maintaining a faithful inference speed. Experiments on point cloud classification show that we achieved state-of-the-art accuracy on ModelNet40 and ScanObjectNN datasets. We also generalize our model to point cloud part segmentation and demonstrate competitive performance on the ShapeNetPart dataset. Code is available at https://github.com/ShenZheng2000/PointNorm-for-Point-Cloud-Analysis.
翻译:由于点云数据结构的不规则性,点云分析具有挑战性。现有的工程通常使用点Net++的特设取样组群操作。为减少点云的不规则性,我们在取样组群操作后引入了一个双角云模块。双角云模块包括点常态化,使组点与抽样点正常化,以及逆向点常态化,使抽样点与组点正常化。在本文中,我们指出,在非点云的典型取样组作业给随后的 MLP 层造成了学习困难。为减少点云的不规则性,我们在取样组作业后引入了一个双角云模块。双角模型包括点常态化,使组点与抽样点正常化,使过去几年的样本点常态和性能饱和性。拟议的点利用地方平均值和全球标准偏差,从2000年本地和全球特征中获益,同时保持忠实的精确度速度。关于点云层分类的实验显示,我们在模型/轨道网段实现了状态模型的准确性能标点,并在ScampOnB部分展示了我们通用的Sqamp Sqoal 数据。