结构感知的自监督点云上采样网络  

Structure-Aware Self-Supervised Point Cloud Upsampling Network

作　　者：徐天 张长伦 

机构地区：[1]北京建筑大学理学院,北京

出　　处：《人工智能与机器人研究》2025年第2期279-289,共11页Artificial Intelligence and Robotics Research

基　　金：国家自然科学基金(No. 62072024)。

摘　　要：现有的多分支自监督点云上采样框架的设计机制通常没有关注到各分支间的结构一致性问题,特征表示能力不足。为解决上述问题,提出一种结构感知的自监督点云上采样网络。特征提取阶段,模型在低层特征处加入结构感知模块,重新组合特征点云,来提升结构感知的能力。同时添加自注意力机制,整合不同分支的特征。特征扩展阶段,提出使用插值点云来指导上采样过程,以提高保持原始点云形状的能力。在PUGAN数据集上进行实验,与基线模型SPU-Net相比,所提出模型在三个指标上均有所提升,其中P2F和HD两个指标提升显著。实验结果表明,模型可以生成具有较多几何细节、较高质量的密集点云。The design mechanisms of existing multi-branch self-supervised point cloud upsampling frameworks usually do not pay attention to the structural consistency problem among branches, and the feature representation capability is insufficient. To solve the above problem, a structure-aware self-supervised point cloud upsampling network is proposed. In the feature extraction stage, the model adds a structure-aware module at the low-level features to recombine the feature point cloud to improve the ability of structure-awareness. At the same time, a self-attention mechanism is added to integrate features from different branches. In the feature expansion stage, the use of interpolated point clouds is proposed to guide the upsampling process to improve the ability to maintain the shape of the original point cloud. Experiments are conducted on the PUGAN dataset. Compared to the baseline model SPU-Net, the proposed model improved on all three metrics. Two of the metrics, P2F and HD, improved significantly. Experimental results show that the model can generate dense point clouds with more geometric details and higher quality.

关 键 词：深度学习 点云上采样 自监督 结构感知 插值引导 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]