摘要
点云分类作为三维视觉领域的核心任务,面临特征表达能力有限与排列不变性处理不足的双重挑战。针对传统基于多层感知机(MLP)的网络难以有效捕捉全局特征及动态聚合局部信息的问题,文章提出一种基于Mamba与循环最大池化的双流增强式点云分类网络。首先,通过引入Mamba模块对原始点云进行序列化建模,利用其长程依赖捕捉能力提取具有强区分性的全局特征;其次,加入循环最大池化(RMP)模块,通过多级迭代的池化操作显式提取点云排列不变特征,并结合循环机制实现局部特征的动态强化与上下文融合。双流架构中,全局特征与局部特征经自适应加权后输入MLP分类头,完成高阶语义推理。在ModelNet40与ScanObjectNN基准数据集上的实验表明,本文方法的分类准确率分别达到93.9%与86.8%,都高于先进的分类方法。消融实验进一步验证了Mamba的全局建模能力与RMP模块对无序点云的鲁棒性增强效果。
Point cloud classification,as a core task in the field of 3D vision,is faced with the dual challenges of limited feature expressiveness and insufficient handling of alignment invariance.Aiming at the problem that traditional multilayer perceptron(MLP)-based networks can hardly effectively cap-ture global features and dynamically aggregate local information,this paper proposes a dual-stream augmented point cloud classification network based on Mamba and cyclic max-pooling.First,the original point cloud is serialized and modeled by introducing the Mamba module,and its long-range dependency capture ability is used to extract global features with strong discriminative prop-erties;second,the Recurrent Maximum Pooling(RMP)module is added to explicitly extract the point cloud arrangement-invariant features through multi-level iterative pooling operations,and combined with the recurrent mechanism to achieve dynamic enhancement and contextual fusion of local features.In the dual-stream architecture,global features and local features are adaptively weighted and input into the MLP classification header to complete the higher-order semantic infer-ence.Experiments on ModelNet40 and ScanObjectNN benchmark datasets show that the classifica-tion accuracy of this paper’s method reaches 93.9%and 86.8%,respectively,both of which are higher than the state-of-the-art classification methods.The ablation experiments further validate the global modeling capability of Mamba with the robustness enhancement effect of the RMP mod-ule on disordered point clouds.
作者
柴国强
Guoqiang Chai(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)
出处
《建模与仿真》
2025年第5期503-515,共13页
Modeling and Simulation
关键词
点云分类
深度学习
三维视觉
特征提取
Point Cloud Classification
Deep Learning
3D Vision
Feature Extraction
