3D Point Cloud Generation Based On Generative Adversarial Networks

Point cloud is a mainstream representation of 3D data with high accuracy and density,and it has extensive applications in fields such as robot vision,autonomous driving,and building model reconstruction.Generating 3D point cloud data is an open research field aimed at promoting the learning of point representations on non-Euclidean objects,which helps to address computer vision tasks such as segmentation,classification,object detection,and feature extraction.In recent years,point cloud generation models based on Generative Adversarial Networks(GANs)has been widely studied.Unlike 2D images,point cloud generation needs to represent discrete points in 3D space without the regular structure.Traditional GANs sample the initial latent code from a Gaussian distribution,more points will aggregate at junction of different semantic parts without regularization methods,leading to non-uniform point clouds.The non-uniformity causes points to cluster in central area.Other areas are sparse or even hollow.Traditional 3D GANs could not represent accurate probability distribution information during training,and it is inefficient in training and difficult to control the generated results.This thesis studies point cloud generation based on GANs,and the main contributions are as follows:(1)A P-Tree GAN is proposed to address the non-uniformity and low model training efficiency in generating point clouds with Tree GAN.The method efficiently learns the probability distribution information of the point cloud during the generator training.The improved mapping network can optimize the point cloud generation.It also result in a more uniform point clouds with a more structured shape.Latent code in the mapping network can also used to change the local features of the generated point cloud.According to the experimental results on the Shape Net Part dataset,the method achieves 0.098 and 2.231 in JSD(Jensen-Shannon Divergence)and FPD(Fréchet point cloud distance)metrics for the overall 16 classes,respectively,outperforming previous methods in point cloud generation.Compared with traditional 3D point cloud generation models,our results are more uniform,and the model training is more efficient.(2)A style-editing-based point cloud generation method P-Tree GAN2 is proposed to control the point cloud generation task.In the method,the mapping network is improved by using more number of fully connected layers.It enhance the ability of the latent code to represent point cloud features.The improved affine transformation module use up-sampling layer to improve the representation of model.A new latent space,W_k space,is proposed in the mapping network,which makes the generated point clouds easy to edit,such as Style Mix.To perform semantic editing in the W_k-space,we train SVM classifier to find a hyperplane which normal vector is used as the separation boundary.The latent code is transformed across the boundary to change semantic features of the point cloud.According to the experimental results on the Shape Net Part dataset,the P-Tree GAN2 achieves better performance in JSD and FPD metrics for airplane,chair,and overall 16 classes than previous methods.In addition,the visualization results of the experiment demonstrate the good effect of the method in semantic editing in W_kspace.After 6-fold cross validation,the average accuracy of SVM classification of hidden codes in the W_k space reached 73.67%,exceeding the classification accuracy in the Z space and the W space,indicating that hidden codes in the W_k space have better linear separability.