【论文阅读】8-Non-local Scan Consolidation for 3D Urban Scenes

1、重要思想
2、算法步骤
- 2.1、Get repeated instance
- 2.2、Instance --- plane
- 2.3、Plane cluster
- 2.4、Get representative plane( for cluster)
- 2.5、Point project ---remove point
- 2.6、Plane ----line ---region ( remove the points ---denoise)
- - 2.6.1、Plane—line:
  - 2.6.2、Line- cluster :
  - 2.6.3、denoise

1、重要思想

本文要点：基于重复结构的检测与变换，实现3D model 的合并与巩固
Non-local filtering :

Apply to denoise ( remove points)
Key: average similar neighborhoods irrespective of their spatial proximity
算法步骤中2-6均为 denoise!!!

Noise 来源：outliers \ varying sampling density

2、算法步骤

2.1、Get repeated instance

Using user guidance repeated facade elements are extracted as instances.

方法：

Description – encoding – configuration ( using (rigid invariant) relative coordinates
at some pivot point of configuration.)
-- to finding similar configuration of descriptor types
Relative coordinate --- vote for absolute coordinates

Descriptor:
4x4x4 cube
0/1： point number vs threshold
matches: bitwise && Di AND Dj > threshold

2.2、Instance — plane

Each instance is segmented into a set of planes.

原因：Our scans, originating from urban buildings, largely consist of flat faces.

平面拟合时，采用ransac

Plane weight / confidence:
W = w1*w2*w3
W1: area 面积
W2: 同质性—平面点密度的变化（ps: 点密度计算：K-nearest）
W3: 各向异性 —平面点集协方差矩阵特征值比例

本文中ICP的作用：

Repetitive instance  -Planes –alignment 重复单元对应平面的配准！！
采用weighted ICP ( point - confidence)In each ICP iteration,
we subsample the point cloud and associate each point with
its nearest point in the base based on the L2 deviation of their normal
vectors.

2.3、Plane cluster

The instances are registered to a consistent coordinate system, and corresponding planes across instances are identified using parameter space clustering.
Plane space : (nx, ny, d)
Plane measure : L2 norm

方法：按照plane confidence 排序，incrementally add cluster ( cluster R: user assign, according to the data quality)

2.4、Get representative plane( for cluster)

Representative planes are selected for groups of clustered ones, and off-plane denoising performed.

Weighted-L1-media：

权重快速衰减因子，权重—plane density:

ps:( outliers: ----average & L1-media 不行)

2.5、Point project —remove point

Point—project –representative plane
Off-plane：  remove
In- plane: keep

2.6、Plane ----line —region ( remove the points —denoise)

Using a similar scheme in 2D, occupied polygonal regions are identified in respective planes in the in-plane denoising phase.

2.6.1、Plane—line:

Plane—boundary point

根据点的局部领域的分布： 点邻域 协方差矩阵 特征值 比例 vs 阈值

Boundary point- line

运用ransac
Line weight: local point density && length
Line—碎片化：由于remove unsampled points

2.6.2、Line- cluster :

将correspondence plane 的lines 进行聚类，聚类方式同 plane cluster( order\incrementally)
得到representative line（weight-L1-media）representative line—regularize: !! prefer orthogonal arrangement ( 正交排列)

2.6.3、denoise

Out-liers: remove---一般处于sparse region or transparent element( 透明的物体)
In -lier: up-sample

【论文阅读】8-Non-local Scan Consolidation for 3D Urban Scenes相关推荐

论文阅读 [TPAMI-2022] Locally Connected Network for Monocular 3D Human Pose Estimation
论文阅读 [TPAMI-2022] Locally Connected Network for Monocular 3D Human Pose Estimation 论文搜索(studyai.com) ...
论文阅读：Multi-view Convolutional Neural Networks for 3D Shape Recognition
Preface 今天阅读一篇ICCV2015的论文:<Multi-view Convolutional Neural Networks for 3D Shape Recognition>, ...
【论文阅读】Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud
最近在看3d目标检测的文章,感觉看完东西还是要记一下自己的想法,尤其是有思考的文章. 论文题目:Point-GNN: Graph Neural Network for 3D Object Detect ...
论文阅读：Your Local GAN: Designing Two Dimensional Local Attention Mechanisms for Generative Models
1.概要文章主要两点贡献,首先引入了一个新的局部稀疏注意层(local sparse attention layer),该层保留了二维几何形状和局部性. 文章证明,用文章的结构替换SAGAN的密集注 ...
论文阅读——A Comprehensive Study on Deep Learning-Based 3D Hand Pose Estimation Methods综述阅读2
3D手势姿态估计综述本文通过对大量有代表性的论文研究,提出一种基于输入数据模式的新分类法,即RGB.深度或多模态信息.最后,我们展示了在最流行的RGB和基于深度的数据集上的结果,并讨论了这一快速 ...
【论文阅读】Globally Consistent and Tightly Coupled 3D LiDAR Inertial Mapping
这篇论文提出了一个改进的地图生成方法,使用的传感器是激光雷达,所以这里也划分到激光SLAM的范畴下了.论文提出的框架的改进主要在两方面,一方面是使用了基于体素化的GICP匹配成本因子,采用这种方法可以 ...
论文阅读笔记：(2021.10 CoRL) DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries
论文地址:DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries | OpenReviewWe introduc ...
论文阅读：Semantic Mapping for Mobile Robots in Indoor Scenes: A Survey
笔记纲要 Basic Information Main Body Introduction Definitions of Semantic Map Spatial Mapping Acquisitio ...
论文阅读-----使用可分离脚印的x射线3D CT向前和向后投影
Long Y , Fessler J A , Balter J M . 3D Forward and Back-Projection for X-Ray CT Using Separable Foot ...
论文阅读笔记：(2021.06, cvpr) Monocular 3D Object Detection: An Extrinsic Parameter Free Approach
这是一篇发表在了cvpr 2021上,能够在线估计外参的单目3D目标检测算法,借鉴了visual odometry和style transfer的方法,效果好, 速度快(~30ms), 意料之外,情理 ...

【论文阅读】8-Non-local Scan Consolidation for 3D Urban Scenes

【论文阅读】8-Non-local Scan Consolidation for 3D Urban Scenes

1、重要思想

2、算法步骤

2.1、Get repeated instance

2.2、Instance — plane

2.3、Plane cluster

2.4、Get representative plane( for cluster)

2.5、Point project —remove point

2.6、Plane ----line —region ( remove the points —denoise)

2.6.1、Plane—line:

2.6.2、Line- cluster :

2.6.3、denoise

【论文阅读】8-Non-local Scan Consolidation for 3D Urban Scenes相关推荐

最新文章

热门文章