目前,基于深度学习的点云上采样方法缺失对局部区域特征关联性的关注和对全局特征的多尺度提取,导致输出的密集点云存在异常值过多、细粒度不高等问题.为解决上述问题,提出了嵌入注意力机制的并行多尺度点云上采样网络(Parallel Multi-scale with Attention mechanism for Point cloud Upsampling),网络由特征提取器、特征拓展器、坐标细化器和坐标重建器4个模块级联组成.首先给定一个N×3的稀疏点云作为输入,为了获得点云的全局和局部特征信息,设计了一个嵌入注意力机制的并行多尺度特征提取模块(PMA)用于将三维空间的点云映射到高维特征空间.其次使用边缘卷积特征拓展器拓展点云特征维度,得到高维点云特征,以更好地保留点云特征的边缘信息,将高维点云特征通过坐标重建器转换回三维空间中.最后使用坐标细化器精细调整输出点云细节.在合成数据集PU1K上的对比实验结果表明,PMA-PU生成的密集点云在倒角距离(CD)、豪斯多夫距离(HD)和点面距离(P2F)上都有显著提升,分别比性能次优的网络模型优化了 7.863％,21.631％,14.686％.可视化结果证明了 PMA-PU具有性能更好的特征提取器,能够生成细粒度更高、形状更接近真实值的密集点云.
Parallel Multi-scale with Attention Mechanism for Point Cloud Upsampling
The current deep learning-based point cloud upsampling method lacks the attention to a local area feature correlation and multi-scale extraction of global features,resulting in the dense output point cloud with too many outliers and low fine-grained granularity.To solve the above problem,a parallel multi-scale with attention mechanism for point cloud upsampling(PMA-PU)network is proposed,which consists of a feature extractor,a feature expander,a coordinate refiner and a coordinate reconstructor.Firstly,giving an N× 3 sparse point cloud as input,a parallel multi-scale feature extraction module(PMA)with an embedded at-tention mechanism is designed to map the point cloud in 3D space to the high-dimensional feature space to obtain the global and local feature information of the point cloud.Secondly,the high-dimensional point cloud features are obtained after expanding the dimensionality of the point cloud features using the edge convolution feature expander to better preserve the edge information of the point cloud features,and the high-dimensional point cloud features are converted back to the 3D space by the coordinate re-constructors.Finally,the output point cloud details are fine-tuned by using the coordinate refiners.The results of the PMA-PU comparison experiments on the synthetic dataset PU1K show that the generated dense point cloud has significant improvement in the three evaluation metrics,Chamfer Distance(CD),Hausdorff Distance(HD),and P2F(point-to-surface),which are significantly better than the second highest performance.The network models with the second highest performance are optimized by 7.863％,21.631％,and 14.686％,respectively.The visualization results demonstrate that PMA-PU has a better performce feature extrac-tor,which can generate dense point clouds with higher fine granularity and a shape closer to the true value.
3D point cloudDeep learningPoint cloud upsamplingParallel multi-scale feature extractionAttention mechanism
万方数据
维普
