意图注意力引导的小样本3D点云目标检测  

Few-shot 3D Point Cloud Object Detection Using Intention-Attention Guidance

作　　者：徐守坤[1] 张路军 石林[1] 刘毅 XU Shoukun;ZHANG Lujun;SHI Lin;LIU Yi(School of Computer and Artificial Intelligence,Aliyun School of Big Data,School of Software,Changzhou University,Changzhou 213164,Jiangsu,China)

机构地区：[1]常州大学计算机与人工智能学院、阿里云大数据学院、软件学院,江苏常州213164

出　　处：《计算机工程》2024年第12期288-295,共8页Computer Engineering

基　　金：国家自然科学基金青年项目(62001341);江苏省自然科学基金面上项目(BK20221379)。

摘　　要：现有的点云目标检测方法往往要求严苛的监督式数据集,这带来了人力、物力等方面的挑战。因此提出一种创新性的解决方案,即采用元学习框架来克服大量标注数据的困扰。对小样本学习技术在3D点云目标检测中的应用进行研究。这一方法能够基于有限的新类别标注样本,预测未标注样本的分类,从而在有限数据条件下仍能取得良好效果。引入原型投票网络来学习类别的几何原型以及支持集的类别原型。此外,为了学习点云上下文信息,引入意图注意力机制,以实现更加精准的信息融合。在原型生成方面,为避免点云原型过度依赖最大池化而丧失大量信息,采用平均池化方法来生成原型。与基准数据集上的基线模型相比,该方法呈现出显著且一致的提升效果,为点云目标检测领域的研究和应用提供了有力支持。Current point-cloud object detection methods often rely on supervised datasets,which pose challenges related to work force and resources.To address this,an innovative meta-learning framework is proposed to reduce the dependency on large labeled datasets.This research explores applying few-shot learning techniques to 3D point-cloud object detection,enabling the classification of unlabeled samples using only a few labeled examples from new classes,achieving strong performance under limited data conditions.A prototypical VoteNet is introduced to learn geometric prototypes of categories and support set prototypes.An intention-attention mechanism is also employed to capture point-cloud contextual information for more precise information fusion.Mean pooling is applied to mitigate overreliance on max-pooling,preventing the loss of critical information during prototype generation.Compared with baseline models on benchmark datasets,the proposed method consistently demonstrated significant improvements,underscoring its potential for further research and practical applications in point-cloud object detection.

关 键 词：点云 目标检测 小样本 投票网络 意图注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]