Cascaded Refinement Network for Point Cloud Completion

Abstract

Point clouds are often sparse and incomplete. Existing shape completion methods are incapable of generating details of objects or learning the complex point distributions. To this end, we propose a cascaded refinement network together with a coarse-to-fine strategy to synthesize the detailed object shapes. Considering the local details of partial input with the global shape information together, we can preserve the existing details in the incomplete point set and generate the missing parts with high fidelity. We also design a patch discriminator that guarantees every local area has the same pattern with the ground truth to learn the complicated point distribution. Quantitative and qualitative experiments on different datasets show that our method achieves superior results compared to existing state-of-the-art approaches on the 3D point cloud completion task. Our source code is available at https: // github.com/ xiaogangw/ cascaded-point-completion.git.

ASFM-Net: Asymmetrical Siamese Feature Matching Network for Point Completion

“…Zhang et al [62] propose a feature aggregation strategy to preserve the primitive details. Wang et al [48] design a cascaded refinement and add the partial input into the decoder directly for high fidelity. However, these methods use global features only extracted from partial inputs, leading to information loss in the encoding process.…”

“…Although the asymmetrical Siamese auto-encoder can extract a more effective global feature and generate a coarse point cloud 𝑃 𝑐𝑜𝑎𝑟𝑠𝑒 , the fine details of the input are inevitably lost. To preserve the detailed information of the input point cloud, following [48], we concatenate the partial inputs with the 𝑃 𝑐𝑜𝑎𝑟𝑠𝑒 to form a synthetic point cloud 𝑃 𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐 using the farthest points sampling (FPS) algorithm and mirror operations. We explore various symmetry operations, including plane symmetry, projective symmetry, and affine transformation operations.…”

3D Point Cloud Restoration via Deep Learning: A Comprehensive Survey

“…PointNet 模型 [12] 首先将深度学习技术引入点 云处理中, 以一组无序点作为输入自动学习用于 对象分类和分割的特征, 并验证了它能够用于点 云数据的多种认知任务, 如分类、语义分割和目标 识别等. 由此, PointNet 或其部分结构或其扩展模 块常常应用到三维点云修复模型中作为特征提取 或特征提取的一部分 [10,20,34] PointNet++模型 [13] 能 够在不同尺度提取局部特征, 通过多层网络结构 得到深层特征, 它的多层次特征提取技术也常常 被应用到点云修复模型中 [20,[35][36]41] . GCN 是一种神经网络架构, 它可以利用图的 结构, 以卷积的方式从邻域聚集节点信息, 如图 14 所示.…”

“…点云修复质量常用的评价函数包括均值 [41] 、均 方根 [41] 和 MSE [29] , CD [56] , EMD [57] , NUC(normalized uniformity coefficients) [20] , P2F(point-to-surface) [11,18] , FPD(Fréchet point cloud distance) [34] , RMSD(root mean square distance-to-surface) [40] , C2C(cloud-tocloud), HD 等. 其中, 集合 A 到集合 B 的 HD 是一 个极大极小函数, 定义为 HD( , ) max{min{ ( , )}}.…”

3D Grid Transformation Network For Point Cloud Completion

“…The foldingbased method transforms the 2D grid points to the surface points of the 3D point clouds by using Multi-Layer Perceptions (MLPs). Due to the achieved success of PCN and folding-based method, many researchers [14,15] follow the idea to utilize 2D grids to generate surfaces for point cloud completion. However, using MLPs directly to transform 2D grid points will retain the original 2D grid structure.…”