结合MAML和Dirichlet过程的小样本点云分类  

Few-Shot Point Clouds Classification Based on MAML and Dirichlet Process

作　　者：刘复昌 李晨璇 王延斌 缪永伟 Liu Fuchang;Li Chenxuan;Wang Yanbin;Miao Yongwei(School of Information Science and Technology,Hangzhou Normal University,Hangzhou 311121;College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016)

机构地区：[1]杭州师范大学信息科学与技术学院,杭州311121 [2]南京航空航天大学机电学院,南京210016

出　　处：《计算机辅助设计与图形学学报》2023年第11期1674-1682,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61972458);浙江省自然科学基金(LZ23F020002,LY20F020017)。

摘　　要：点云被广泛使用在各种三维应用场景中,但是实际应用中通常存在扫描、标注费时费力等局限性,因此基于小样本数据集的点云分类网络更加符合应用需求.为了有效地提高深度学习分类算法在小样本点云数据集上的分类效果,提出一种针对小样本数据集的点云分类方法.针对训练数据集不平衡问题,首先采用基于相似度依赖的Dirichlet中餐馆过程对数据集进行预处理,在无需人工指定聚类个数的前提下对样本进行重新聚类,以提升分类网络在小样本数据集上的性能;然后在重新聚类后的样本上使用模型无关(model agnostic meta learning,MAML)算法训练PointNet++,达到用少量点云样本就能快速适应新任务的能力.所提方法不但降低了模型对数据量的依赖,提高了模型泛化能力,而且成功地把MAML算法从二维图像分类拓展到三维点云分类中;在Modelnet40数据集上的实验结果表明,与PointNet++相比,该方法的训练时间减少了一半,分类准确率平均提高6.67%,验证了该方法在小样本数据集上的有效性.Point clouds are widely used in various 3D application scenes.However,there are some issues in the real-world scenarios,such as labor-consuming and time-costing on data annotation.Few-shot point clouds classification neural networks are popular and can satisfy practical application requirements.To improve the performance of deep learning models in small sample point cloud classification,we propose a point clouds classification method based on MAML and Dirichlet process.To handle the imbalance of data sets,this paper proposes a similarity dependent Dirichlet process called Chinese restaurant model to cluster the dataset.This cluster process can complete clustering without manually specifying the number of clusters and result in higher classification accuracy with the optimal clustering.Subsequently,we apply MAML framework to train Point-Net++network based on samples processed by our cluster model.With the help of our data clustering,the MAML algorithm can adapt to new tasks quickly through a small number of data samples.The proposed method alleviates the issue of data hungry on deep learning for point clouds and improve the generalization of neural networks,as well as successfully extending the MAML algorithm from two-dimensional image classification to three-dimensional point clouds classification.Compared with PointNet++model,the training time is reduced by half and the overall accuracy is improved by 6.67%.Experimental results demonstrate the effectiveness of our proposed small sample point cloud classification algorithm on public dataset.

关 键 词：小样本学习 点云分类 模型无关的元学习 中餐馆过程 Dirichlet过程 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]