| 摘 要: 受点云非结构化、无序性等特性的影响,一些现有的自注意力方法不能充分提取上下文语意特征,基于此提出了一种用于点云特征提取的局部邻域Transformer(Local Neighborhood Transformer,LNT)。首先,通过最远点采样(FPS)和K最近邻算法(KNN)对点云进行邻域划分。其次,结合相对位置编码,在各个邻域内计算局部自注意力,达到线性计算复杂度。最后,通过连接操作以及线性层捕获点云局部特征。此外,设计了点云多特征融合方法对各层特征信息进行聚合,以提高模型的性能。实验结果表明:该方法在ModelNet40数据集中分类的总体精度可达到93.3%,比PCT提升了0.1%;类平均精度可达到92.0%,比PointMLP提升了0.6%。同时,在ShapeNet数据集中的点云分割结果也是有效的。 |
| 关键词: 点云处理;Transformer;特征融合;神经网络 |
|
中图分类号: TP183
文献标识码: A
|
|
| Point Cloud Feature Extraction Method Based on Local Neighborhood Transformer |
|
ZHANG Haibo1, SHEN Yang1,2, XU Hao2, BAO Yanxia2,3, LIU Jiang2
|
(1.School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.School of Engineering, Lishui University, Lishui 323000, China;
3.Zhejiang Zhangxin Media Technology Co., Ltd., Lishui 323020, China)
linhai992@163.com; 1178409849@qq.com; oah_ux@126.com; 82240849@qq.com; elecliu@lsu.edu.cn
|
| Abstract: Affected by the unstructured and disordered nature of point clouds, some existing self-attentive methods cannot fully extract contextual semantic features, based on which a Local Neighborhood Transformer (LNT) for point cloud feature extraction is proposed. Firstly, neighborhood partitioning of point clouds is achieved through Farthest Point Sampling (FPS) and K-nearest Neighbor (KNN) algorithm. Secondly, local self-attention is computed within each neighborhood by combining relative position encoding to achieve linear computational complexity. Finally, the local features of the point cloud are captured by the join operation as well as the linear layer. In addition, a point cloud multi-featured fusion method is designed to aggregate the feature information of each layer to improve the performance of the model. The experimental results show that the overall accuracy of the method can reach 93.3% in the ModelNet40 dataset for classification, which is 0.1% higher than PCT; the class average accuracy can reach 92.0% , which is 0.6% higher than PointMLP. Meanwhile, the point cloud segmentation results are also valid in the ShapeNet dataset. |
| Keywords: point cloud processing; Transformer; feature fusion; neural network |
