基于PointNet++改进的点云特征提取与分类网络架构

Improved Point Cloud Feature Extraction and Classification Network Architecture Based on PointNet++

点云作为一种能提供丰富空间信息与几何特征的数据表达形式,正受到越来越多的重视。为了克服其无序性,以及不均匀的空间分布带来的影响,许多研究者采取平面投影,或者使用体素网格对原始点云进行转换,但这些方式都是以损失三维信息或者增大数据规模为前提。PointNet[1]创新性的使用原始点云作为输入,提取特征并处理。PointNet++[2]则在此之上,更进一步地加强了对局部特征的提取能力。论文结合深度学习中的优化思想,对PointNet++结构进行改进,加入自顶向下的网络分支,经过处理后将原网络每一层的中间特征都输入到最终分类网络,更进一步地强化特征提取的能力。该网络易于理解且高效,在ModelNet40数据集上测试,整体分类准确率有明显提升,证明了其优化后的特征提取能力。

As a data representation that can provide rich spatial information and geometric features,point cloud is receiving more and more attention.In order to overcome the effects of disorder and uneven spatial distribution,many researchers use planar projection or voxel grid to transform the original point cloud,but these methods will lose three-dimensional information or increase the scale of the data.PointNet creatively uses the original point cloud as input to extract features and process them.PointNet++is on top of this,further enhancing the ability to extract local features.In this paper,the optimization idea is adopted from deep learning and the PointNet++structure is improved by adding the top-down network branch.After processing,the intermediate features of each layer extracted from the original network are sent into the final classification network.The new branch,according to the experiments,further enhances the ability of feature extraction.The network is easy to understand and very efficient.It is tested on the ModelNet40 dataset,and the overall classification accuracy is significantly improved,which proves its optimized feature extraction ability.