NormalF-Net: Normal Filtering Neural Network for Feature-preserving Mesh Denoising

Li, Zhiqi; Zhang, Yingkui; Feng, Yidan; Xie, Xingyu; Wang, Qiong; Wei, Mingqiang; Heng, Pheng‐Ann

doi:10.1016/j.cad.2020.102861

Cited by 24 publications

(28 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Also [32] propose a learning framework based on a nonlocal similarity approach: Patch vectors based on a similarity criterion are grouped and fed into a convolution network. In contrast, our convolutions have a spatial support, and can extract meaningful local features at different scales.…”

Section: Data-driven Methodsmentioning

confidence: 99%

“…In order to evaluate the benefit of our end-to-end learning architecture, we first compare to current state-of-the-art learning-based approaches for mesh denoising which are the Cascaded Normal Regression (CNR) method of [2] and NormalF-Net [32]. Comparisons on synthetic and real data are presented in sections 7.2 and 7.3 respectively.…”

Section: Evaluation Strategymentioning

confidence: 99%

“…The vertex updating step is a process that can alter the raw measurement, as will be shown in our experiments. However, we note that [2] and [32] while also using the angular difference with respect to ground truth normals, consider the refined normals after the vertex updating step. This seems legitimate in their case since they follow an iterative scheme where vertex updating is an intrisic part of the method.…”

Section: Evaluation Strategymentioning

confidence: 99%

“…BMD [9], BNF [15], GMNF [16], [21] L 0 Minimization, bayesian method [3], CNR [2], NormalF-Net [32], NLLR [30], Ours (raw normals estimated by our network), Ours (final result). [32] for other learning methods are also reported. In our case, the preprocessing step is performed on the CPU only, and could certainly be accelerated with a proper parallel implementation.…”

Section: Runtime Experimentsmentioning

confidence: 99%

“…On the other hand, our method is significantly faster than NormalF-Net [32], and several orders of magnitude faster than NormalNet [33], where a local support is computed for each facet independently. This validates our motivation for using graph convolutions on the mesh.…”

Section: Runtime Experimentsmentioning

confidence: 99%

See 4 more Smart Citations

Mesh Denoising With Facet Graph Convolutions

Armando

Franco

Boyer

2022

IEEE Trans. Visual. Comput. Graphics

View full text Add to dashboard Cite

Section: Data-driven Methodsmentioning

confidence: 99%

Section: Evaluation Strategymentioning

confidence: 99%

Section: Evaluation Strategymentioning

confidence: 99%

Section: Runtime Experimentsmentioning

confidence: 99%

Section: Runtime Experimentsmentioning

confidence: 99%

See 3 more Smart Citations

Mesh Denoising With Facet Graph Convolutions

Armando

Franco

Boyer

2022

IEEE Trans. Visual. Comput. Graphics

View full text Add to dashboard Cite

Learning Self-prior for Mesh Denoising Using Dual Graph Convolutional Networks

Hattori

Yatagawa

Ohtake

et al. 2022

Lecture Notes in Computer Science

View full text Add to dashboard Cite

This study presents a self-prior-based mesh inpainting framework that requires only an incomplete mesh as input, without the need for any training datasets. Additionally, our method maintains the polygonal mesh format throughout the inpainting process without converting the shape format to an intermediate, such as a voxel grid, a point cloud, or an implicit function, which are typically considered easier for deep neural networks to process. To achieve this goal, we introduce two graph convolutional networks (GCNs): single-resolution GCN (SGCN) and multi-resolution GCN (MGCN), both trained in a self-supervised manner. Our approach refines a watertight mesh obtained from the initial hole filling to generate a completed output mesh. Specifically, we train the GCNs to deform an oversmoothed version of the input mesh into the expected completed shape. To supervise the GCNs for accurate vertex displacements, despite the unknown correct displacements at real holes, we utilize multiple sets of meshes with several connected regions marked as fake holes. The correct displacements are known for vertices in these fake holes, enabling network training with loss functions that assess the accuracy of displacement vectors estimated by the GCNs. We demonstrate that our method outperforms traditional dataset-independent approaches and exhibits greater robustness compared to other deep-learning-based methods for shapes that less frequently appear in shape datasets.

show abstract