5. Post‐processing

Repairing, denoising, smoothing, simplification…

Repairing (completion, hole‐filling, restoration)

(1) Filling Holes in Meshes 

[Liepa, SGP 2003]

Pipeline

  1. Hole identification
  2. Hole triangulation
  3. Mesh refinement 
  4. Mesh fairing

对空洞部分三角化 ⇒ 三角形加密

Summary

  • Easy to implement
  • Focus algorithm on holes
  • Triangulation may self‐intersect
  • Can’t fill holes with islands
  • Fairing weaken original surface feature

(2) Robust Repair of Polygonal Models [Ju, Siggraph 2004]

没讲，能保证拓扑关系

  • Employ a space‐efficient octree grid
  • Produce closed, manifold surface for any input  model

(3) Context‐based Surface Completion

[Sharf et al., Siggraph 2004]

Motivation

Complete the missing region with patches that conform with its context 

学习周边的几何细节，再插值

Method

Import patches with matching context from the surface itself:

  • Analyze surface characteristics.
  • Find best matching patch.
  • Fit imported patch to boundary.

Limitations: Semantics

(4) Example‐based Surface Completion

[Pauly et al., SGP 2005]

  • Use 3D model database to provide geometric priors for shape completion
  • Apply non‐rigid transforms on the models
    • More deformation \(\Rightarrow \) less likely completion
  • Consistently combine geometric information from multiple context models
  • Final result comes with confidence values

从数据库里相似的数据来修复缺失的数据

(5) Atomic Volumes for Mesh Completion 

[ Podolak and Rusinkiewicz, SGP 2005]

Atomic Volumes

A volume is atomic if it doesn’t intersect the polygons of the mesh.

利用内部信息

Summary　　　

  • Avoid changing, approximating or re‐sampling the original mesh data
  • Incorporate user constraints
  • Can’t process holes with islands

(6) Geometry Completion by Texture Synthesis

[Nguyen et al., PG 2005]

在图像空间做纹理合成，再恢复到3D空间中

(7) Others

  • Marco Attene, Marcel Campen, Leif Kobbelt. Polygon Mesh Repairing: An Application Perspective. ACM Computing Surveys, 2012.

  • Learning based 3D data completion in recent years

    • Han et al. High Resolution Shape Completion Using Deep Neural Networks for Global Structure and Local Geometry Inference. ICCV 2017.
    • Han et al. Deep Reinforcement Learning of Volume‐guided Progressive View Inpainting for 3D Point Scene Completion from a Single Depth Image. CVPR 2019.
    • Nie et al. Skeleton‐bridged Point Completion: From Global Inference to Local Adjustment. NeuIPS 2020.

本文出自CaterpillarStudyGroup，转载请注明出处。 https://caterpillarstudygroup.github.io/GAMES102_mdbook/