3D点云形状补全GAN

在真实的扫描环境中,由于视线遮挡或技术人员操作不当,实际采集到的点云模型会存在形状不完整的问题。点云模型的不完整性会对后续应用产生严重的影响,因此提出3D点云形状补全GAN用于完成点云模型的形状补全。该网络的点云重建部分由PointNet中用于数据对齐的T-Net结构与3D点云AutoEncoder网络相结合,来完成预测和填充缺失数据,识别器采用3D点云AutoEncoder中的Encoder部分对补全3D点云数据与真实的3D点云数据进行区分。最后,在ShapeNet数据集中训练上述网络结构,对所训练的网络模型进行验证并与其他基准方法进行定性比较。从实验结果可以看出,3D点云形状补全GAN可以将具有缺失数据的点云模型补全为完整的3D点云。在ShapeNet的3个子数据集chair,table以及bed上,相比基于3D点云AutoEncoder的方法,所提方法的F 1分数分别提高了3.0%,3.3%以及3.1%,相比基于体素3D-EPN的方法,所提方法的F 1分数分别提高了9.9%,5.8%以及4.3%。

3D Point Cloud Shape Completion GAN

In the real scanning environment,due to the occlusion of line of sight or improper operation of technicians,the actual point cloud model has incomplete shape.The incompleteness of point cloud model has a serious impact on subsequent applications.Therefore,this paper proposes a 3D point cloud shape completion generative adversarial networks to complete the shape completion of point cloud model.The point cloud reconstruction part of the network combines the T-Net structure used for data alignment in PointNet with the 3D point cloud AutoEncoder network to complete the prediction and fill in the missing data.The discriminator uses the Encoder part of the 3D point cloud AutoEncoder to distinguish the completed 3D point cloud data from the real 3D point cloud data.Finally,in the ShapeNet trained the above network structure,the trained network model is verified and compared with other benchmark methods qualitatively.From the experimental results,it can be seen that the 3D point cloud shape completion generation adversarial network can complete the point cloud model with missing data into a complete 3D point cloud.In ShapeNet’s three sub-datasets chair,table,and bed,compared with the method based on 3D point cloud AutoEncoder,the F 1 score is improved by 3.0%,3.3%and 3.1%,and compared with the method based on the voxel 3D-EPN method,the F 1 score is increased by 9.9%,5.8%,and 4.3%,respectively.