Differentiable Manifold Reconstruction for Point Cloud Denoising

Abstract

3D point clouds are often perturbed by noise due to the inherent limitation of acquisition equipments, which obstructs downstream tasks such as surface reconstruction, rendering and so on. Previous works mostly infer the displacement of noisy points from the underlying surface, which however are not designated to recover the surface explicitly and may lead to sub-optimal denoising results. To this end, we propose to learn the underlying manifold of a noisy point cloud from differentiably subsampled points with trivial noise perturbation and their embedded neighborhood feature, aiming to capture intrinsic structures in point clouds. Specifically, we present an autoencoder-like neural network. The encoder learns both local and non-local feature representations of each point, and then samples points with low noise via an adaptive differentiable pooling operation. Afterwards, the decoder infers the underlying manifold by transforming each sampled point along with the embedded feature of its neighborhood to a local surface centered around the point. By resampling on the reconstructed manifold, we obtain a denoised point cloud. Further, we design an unsupervised training loss, so that our network can be trained in either an unsupervised or supervised fashion. Experiments show that our method significantly outperforms state-of-the-art denoising methods under both synthetic noise and real world noise. The code and data are available at https://github.com/luost26/DMRDenoise. CCS CONCEPTS • Computing methodologies → Point-based models; 3D imaging.

Incomplete Region Estimation and Restoration of 3D Point Cloud Human Face Datasets

“…However, we use these models for restoration purposes. To compare the proposed method with a more recent point cloud processing technique, we use a point cloud denoizing method called differential manifold reconstruction for denoizing (DRMD) [ 33 ] which is more appropriate for the restoration of the estimated point cloud. In this method, Luo et al tried to remove the noise to improve the quality of the noisy point cloud.…”

ChartPointFlow for Topology-Aware 3D Point Cloud Generation

“…Compared with AtlasNets, ChartPointFlow has an architecture that is more consistent with the definition of charts. Luo and Hu [30] also introduced a similar concept for denoising.…”

“…As a representation of an object's surface, a point cloud often has a thin, circular, or hollow structure [30]. Flow-based generative models encounter a difficulty in expressing such manifold-like structures because a bijective map that is necessary for these models does not exist between a Euclidean space and a manifold with holes, as shown in the top panel of Fig.…”

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

“…Zhang 等 [35] ModelNet [64] 合成 规整, 结构简单 PU-Net [20] , 3PU [21] , PU-GAN [18] Pistilli 等 [37] , Duan 等 [39] , Luo 等 [42] Sketchfab 合成 规整, 高分辨率 3PU [21] Paris-rue-Madame [65] 真实 带标注, 不均匀, 有噪声, 不完整 NLPA [19] , Luo 等 [42] SHREC15 [66] 合成 规整, 需采样点集 PU-Net [20] Matterport3D [67] 真实 带标注, 不均匀, 有噪声和空洞 Chen 等 [28] Visionair repository 数据集 ① 中的点云相对规 整, 包含了从光滑的非刚性物体到陡峭的刚性物 体等多种点云模型. 使用者可以根据需要自行选 择三维模型的分辨率.…”

“…Paris-rue-Madame [65] 是来自 rue Madame 的公 共三维标记数据集, 它包含了使用移动激光扫描 系统扫描得到的 160m 长的街道段物体点云, 带有 点三维坐标、反射率、对象标签和对象类. 其类别 包括建筑物、地面、汽车、摩托车、行人和交通标 志等, 可用于基准城市检测、分割和分类等, 在点 云修复中主要用作模型测试 [19,42] . SHREC 是 Eurographics 举办的每年一届的三 维模型检索大赛.…”