摘要
针对当前基于深度学习的点云理解任务需要大量标注数据但数据标注极为消耗成本的现实问题,提出一种基于对比学习预训练的点云主动筛选点云标注方法.通过交替运行对比学习预训练特征提取与主动学习选择模块,在未标注数据中筛选最有代表性的样本进行标注,从而在有限的标注成本下获得最佳性能的点云理解模型.首先基于对比学习的自监督范式进行预训练;然后固定模型参数,利用该模型对未标注点云提取特征,通过设计基于不确定性和特征多样性的指标,从中选择代表性数据进行标注.在点云分类以及分割等任务中,验证了所提方法的有效性;在ModelNet40数据集上的实验结果表明,该方法可有效地提高模型在弱监督下的表现,与随机选择数据进行标注的方法相比,可以提高20%以上的准确率,在接近10%的数据标注下最终达到73%的准确率;在ShapeNet数据集上的实验结果证明,该方法对于分割任务也有较好的表现,在1000组标注数据下取得了91%的精度,接近于监督训练水平.
Aiming at the practical problem that the current point cloud understanding task based on deep learning requires a large amount of labeled data and the labeling of 3D point clouds is extremely time-consuming,we propose an active point cloud labeling method based on contrastive pre-training.The unlabeled samples are actively screened and labeled by alternately running the contrastive learning pre-trained module and the active learning labeling module,thereby improving the model performance under weak supervision.This proposed method first pre-trains the feature extraction module based on the idea of contrastive learning,fixes the parameters of the model,and then uses the model for the active learning module,designs selection indicators based on uncertainty and feature diversity,and calculates the features of unlabeled point cloud data,then selects the desired points to be labelled.The validity of this method is verified in the point clouds understanding tasks.The Model-Net40 dataset is used for verification.The experimental results show that this method can effectively improve the performance of the model under weak supervision.Compared with the randomly selected method,our method can improve the accuracy rate by more than 20%,and finally achieve 73%accuracy under the data labeling,close to 10%on ModelNet.The result on ShapeNet with little data is also promising,which is close to the level of supervised training with the accuracy of 91%under 1000 annotations.
作者
杨国庆
赖文韬
黄惠
Yang Guoqing;Lai Wentao;Huang Hui(Visual Computing Research Center,Shenzhen University,Shenzhen 518060)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2023年第11期1664-1673,共10页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(62161146005,U21B2023)
广东省高等学校创新团队项目(2022KCXTD025)
深圳市科技创新项目(KQTD20210811090044003,RCJC20200714114435012,JCYJ20210324120213036)
深圳大学研究生教育改革项目(SZUGS2022JG01)。
关键词
点云理解
对比学习
主动标注
弱监督训练
point cloud understanding
contrast learning
active learning
weakly-supervised learning
