一种基于注意力机制的三维点云物体识别方法

三维点云数据通常具备无序排列的结构。在三维点云数据处理领域,深度学习模型通常会利用最大池化等对称操作来处理点云的排列不变性。最大池化方法一方面会破坏点云的信息结构,使得局部信息与全局信息难以交互。另一方面,最大池化方法对点云信息过度压缩,得到的特征对局部细节描述不足。针对上述问题,提出了AttentionPointNet的网络结构。该网络利用注意力机制,使每个点与点云其余部分进行特征交互,实现了局部与全局信息的综合。为降低最大池化造成的信息损失,提出了一种稀疏卷积方法来替代池化操作。这种方法利用大步长的稀疏卷积实现全局信息的提取。在ModelNet40数据集上,AttentionPointNet取得了87.2%的准确率。不使用池化层,完全采用卷积层实现的模型取得了86.2%的分类准确率。

A 3D Point Cloud Object Recognition Method Based on Attention Mechanism

3D point cloud data usually has an unordered structure.In the field of point cloud data processing,deep learning models usually use the symmetry operations such as maximum pooling to deal with the permutation invariance of point clouds.On the one hand,this approach often destroys local information of point cloud data.On the other hand,the maxpooling method over-compresses point cloud information,and the extracted features are insufficiently described for local details.Aiming at those problems,we propose a network structure called AttentionPointNet which uses the attention mechanism to make each point interact with the rest of the point cloud to achieve the integration of local and global information.In order to reduce the information loss caused by the maximum pooling,we propose a sparse convolution to replace the pooling layer,which uses large stride sparse convolution to extract global information.On the ModelNet40 dataset,AttentionPointNet achieves 87.2%classification accuracy.The model,which only uses convolution layers to replace maxpooling layer,achieves 86.2%classification accuracy.