Research On Dnn Pooling Operations For 3D Point Cloud

In recent years,with the promotion of deep learning on 3D point cloud data,the research on convolutional neural network models for 3D point cloud data has received extensive attention.However,due to the irregularity topology of the 3D point cloud data structure,another important component that promotes the success of deep learning,the pooling method,is less popular.Pooling operation in the deep learning framework not only plays a role in improving computing efficiency,but also improves the robustness of the model.It is an indispensable part of deep learning in the application of 3D point cloud.The pooling method suitable for 3D point clouds not only needs to provide a complete low-resolution point cloud topological structure for the subsequent convolutional layer,but also needs to retain important points on semantic information and their corresponding features.In response to the above problems,this thesis proposes three pooling methods :(1)A global feature guided pooling algorithm is proposed.Aiming at the problem that the existing methods to construct sparse point clouds only consider the input coordinate information,this thesis proposes a sparse point cloud construction algorithm based on deep learning,which uses global feature to adaptively construct sparse point clouds for downstream neural networks.Aiming at the problem that the sparse point cloud constructed by the adaptive model cannot reflect the complete topological structure of the input point cloud,this paper narrows the distribution of the sparse point cloud and the input point cloud by adding the chamfering distance loss term.Based on the proposed feature pooling algorithm,the multi-layer classification model EC-GDP,on two public object classification datasets,the classification accuracy of the benchmark model is increased by 0.3% and 1.4%,respectively.(2)A random pooling algorithm fusing local features is proposed.From the 3D point cloud classification experiment,it can be seen that the topological structure of the point cloud after pooling has a greater impact on the overall performance of the model.To solve this problem,this paper builds a sparse point cloud topology based on the method of randomly selecting points,and combines the local feature aggregation algorithm to propose a random pooling algorithm.The experimental results on two object classification data sets show that introducing the random pooling algorithm into the single-layer classification model DGCNN can significantly improve the classification accuracy of 0.5% and 1.2%.In addition,for the 3D point cloud segmentation task,this thesis constructs a multi-layer point cloud segmentation model based on the random pooling algorithm,and increases the average m Io U of the benchmark model by 6.1% on the 3D point cloud semantic segmentation data set S3 DIS.(3)A feature pooling algorithm based on mutual information is proposed.The global feature-guided pooling algorithm cannot provide a complete input point cloud topology for the downstream neural network,and the random pooling algorithm cannot adaptively select important features for different tasks.In response to the above two problems,this thesis defines the mutual information between the input point feature and its neighborhood feature to characterize the importance of each point,so that the sparse point cloud topology constructed based on mutual information can retain the topological structure of the input point cloud,It can also retain important feature information for downstream neural networks.Experimental results on two object classification data sets show that the EC-MIP model based on the proposed mutual information pooling algorithm improves the classification accuracy of the benchmark model by 0.6% and2.0%,respectively.