【论文阅读】SurfelMeshing: Online Surfel-Based Mesh Reconstruction

SurfelMeshing: Online Surfel-Based Mesh Reconstruction TPAMI 2018

Abstract

我们解决从实时RGB-D视频重建网格(mesh)的问题，假设一个校准的摄像机和外部提供的姿势（例如，由SLAM系统提供）。与大多数现有的方法不同，我们不在体积(volume)中融合深度测量(fuse depth measurements)，而是在dense surfel cloud中进行融合。我们异步地（重新）三角化平滑的surfels来重建surface mesh。这种新的方法能够在SLAM过程中保持场景的dense surface representation，能够快速适应循环闭合(loop closures)。这可以通过变形(deforming)surfel云并在必要时异步重新划分曲面(remeshing the surface)来实现。基于surfel的表示法自然也支持强变化的扫描分辨率(scan resolution)。特别是，它在输入相机的分辨率下重建颜色。此外，与许多体积法相比，我们的方法可以重建薄物体，因为物体不需要封闭一个体积(since objects do not need to enclose a volume)。

我们在一系列实验中展示了我们的方法，表明它能产生与最新技术相竞争的重建，并讨论了它的优点和局限性。

该算法（不包括循环闭包(loop closure)功能）在https://github.com/puzzlepaint/survelmeshing。

Introduction

比较流行的表示场景的两中方法：

1. voxel volume

   Each voxel in the volume stores the truncated signed distance to the closest surface, which can be updated very efficiently in parallel.
   A continuous surface representa- tion in the form of a mesh can then be extracted from this volume with the Marching Cubes algorithm.
   However, volume-based methods lack flexibility.

2. represent the scene with a set of points or with surfels

   This scene representation is flexible, since point or surfel coordinates can be updated very efficiently for the whole reconstruction.
   It is also highly adaptive as measurements at a higher resolution lead to denser point resp. surfel clouds, and it easily handles thin objects.
   The main drawback of this representation is its discrete nature(离散性), which can be resolved by meshing.

In this paper, we show that online meshing of reconstructed dense surfel clouds is possible through denoising and adaptive remeshing.

SURFEL-BASED MESH RECONSTRUCTION

包含四个部分：

• surfel reconstrction
• surfel denoising
• meshing
• remeshing

其中：
Surfel reconstruction and denoising reconstruct a surfel cloud at the frame rate of the video input.

Meshing and remeshing create and maintain a triangulation of it.

The output is a mesh whose vertices correspond to the surfels, which are updated at frame rate, and whose topology is determined by the triangulation, which is updated at a lower rate

Surfel Reconstruction

Our algorithm triangulates a surfel cloud.

(a) Data associated. 灰色的是新测量的点，黑色的点是在相机视角之外不进行考虑的点，其他颜色的是已经存在的surfels。这里根据模型定义了三种状态：support conflict occlude 。其中绿色的surfels是被新测量的值所support , 红色的surfels代表conflict， 蓝色的代表occlude。

(b) 在整合（a）中的测量值后，conflict surfel被替换(红色), supported surfel用测量值求平均值（绿色），并为没有支持或冲突曲面的测量创建新曲面，即为occlude surfel创建new surfel（灰色）。

© Loops are closed with a non-rigid deformation, aiming to align new surface parts (green) with corresponding old surfaces (blue). 循环以非刚性变形闭合，旨在将新曲面零件（绿色）与相应的旧曲面（蓝色）对齐。

Surfel Denoising

Meshing

中心曲面的三角剖分概述：（a）在曲面的切线平面上进行邻域搜索和投影，如（b）所示，在平面内具有可见性。（c） 在初始三角形创建后。（d） 更新了可见性和邻居排序。（e） 间隙和窄三角形分类。（f） 间隙和窄三角切除后的结果。

Remeshing

Remeshing过程概述：（a）识别出无效三角形（红色）（b） 将删除三角形角点顶点的相邻搜索半径（黄色）内连接到曲面（红色）的所有三角形。受影响的曲面（红色和蓝色）将计划重新划分。（c） 孔由网格算法填充。