基于生成对抗网络的点云形状保结构补全

Structure-preserving shape completion of 3D point clouds with generative adversarial network

针对三维点云形状修复补全中难以保持形状精细结构信息的问题,借助于生成对抗网络框架,本文提出了一种自动修复补全三维点云形状的神经网络结构.该网络由生成器和判别器构成.神经网络的生成器采用编码器–解码器结构,以缺失的三维点云形状作为输入,首先通过输入变换和特征变换对齐输入点云数据的采样点位置与特征信息;然后借助权共享多层感知器对各采样点提取局部形状特征并利用最大池化层与多层感知器编码提取出采样点的特征码字;其次将采样点特征码字加上网格坐标数据,解码器使用2个连续的三层感知器折叠操作将网格数据转变成点云形状的缺失补全数据;最后将缺失补全数据与点云输入数据合并,得到完整的三维点云形状.神经网络的判别器则接收真实的完整点云形状数据和生成器生成的完整点云形状数据,并利用与生成器相同的编码器结构判别出点云形状数据的真假并反馈以不断优化生成器,最终使生成器生成足以"以假乱真"的点云形状数据样本.实验表明,针对形状缺失的稠密点云和稀疏点云数据,本文方法在修复补全形状缺失部分的同时能有效保持输入点云形状的精细结构信息.

Due to the difficulty in maintaining the fine structures of 3 D point cloud in shape completion,this study,with the help of the generative adversarial network framework,proposes a novel neural network for automatically repairing and completing the 3 D shape of point clouds.This network consists of a generator and a discriminator.The generator of the proposed neural network adopts an encoder-decoder structure and takes the missing 3 D point cloud shape data as the input.Firstly,it aligns the sampling point position and feature information of the input point cloud data by the input transform and feature transform.Then the weighted shared multi-layer perceptron extracts the local shape features for each sampling point and also extracts its feature codewords using the maximum pool layer and multi-layer perceptron coding.Secondly,it adds the feature codewords of sampling points with the grid coordinate data,and the decoder converts the grid data into the missing data of the underlying point cloud using two successive three-layer perceptron folding operations.Finally,it merges the missing completion data and the input data to get the complete 3 D point cloud shape.Meanwhile,the proposed neural network discriminator receives the real and the completed point cloud data generated by the generator.The same encoder structure as the generator is also adopted to distinguish the true or false of the point cloud data,while the classification results are a feedback for optimizing the generator.Also,the generator will generate the"real"point cloud shape data.Experimental results illustrate that,for both the dense and sparse incomplete point cloud data,the proposed method effectively maintains the fine structures of the input point clouds while repairing the missing part of the underlying shapes.