融合形状结构恢复和细节补偿的双分支点云修复网络
A Double-Branch Point CloudCompletion Networkby Combining Shape Structure Recoveryand Local Detail Compensation
-
摘要: 针对传统点云修复中难以有效保持原始形状细节结构信息的问题, 提出一种融合形状全局结构恢复和局部细节补偿的双分支修复网络. 网络中的形状全局结构恢复分支为编解码-解码器结构, 编码器对缺失点云数据进行特征变换以克服点云形状的旋转不变性, 利用最大池化操作解决点云的无序性问题, 并通过多层感知器生成原始点云的特征码字, 解码器对编码得到的特征码字使用4个二维网格进行2次折叠操作以拟合点云形状得到粗修复结果; 为了补偿点云粗修复结果的形状细节信息, 网络中的局部细节补偿分支则对编码器提取得到的不同维度特征, 通过层次特征学习和多层次特征融合以学习点云形状的几何结构特征, 从而有效恢复缺失点云数据并保留原始形状细节信息. 最终将经全局结构恢复分支和局部细节补偿分支分别得到的点云数据经拼接融合, 再进行迭代最远点重采样而得到点云形状精修复结果. 实验结果表明, 所提网络在ShapeNet数据集上相比已有网络修复结果的平均CD误差和平均EMD误差分别低16%~29%与19%~65%; 在ModelNet数据集上相比已有网络修复结果的平均CD误差和平均EMD误差分别低6%~41%与31%~59%. 所提出的双分支修复网络可以修复原始形状的整体结构信息, 并能有效恢复其形状细节而生成采样点分布均匀的完整点云模型, 网络对模型噪声和不同程度模型缺失均具有鲁棒性.Abstract: To address the issue that it is difficult to effectively maintain the detail information of original shapes for traditional point cloud restoration methods, a two-branch point cloud completion network is proposed which combines shape structure recovery branch and local detail compensation branch. For the shape structure recovery branch, its encoder performs feature transformation on the missing point cloud data to overcome the rotational invariance of the 3d shapes, and solves the disorder problem of the point cloud by using the maximum pooling operation, which can generate the feature codeword of the input shape by adopting the multi-layer perceptron. In order to compensate shape details of the coarse completion results, the local detail compensation branch can learn the geometric features of the underlying shapes through hierarchical feature learning and multi-level feature fusion for different dimensional features extracted from the encoder, so as to effectively recover the missing point cloud data and retain the original shape details. Finally, the point cloud data obtained by these two branches will be stitched and fused, and then the farthest point resampling is iteratively performed to obtain the final point cloud completion results. Compared with the completion results on ShapeNet dataset, the average CD error and EMD error of our proposed network are 16%~29% and 19%~65% lower than that of the existing networks respectively, whilst the average CD error and EMD error are 6%~41% and 31%~59% lower than that of the existing network respectively if testing on the ModelNet dataset. The proposed two-branch completion network can repair the overall structure of the underlying shape and effectively recover its shape details to generate a complete point cloud model with uniform distribution of sampling points. The network is also robust to model noise and different degrees of missing data.