Consolidation of Unorganized Point Clouds for Surface Reconstruction(WLOP)的代码

论文:: Consolidation of Unorganized Point Clouds for Surface Reconstruction(WLOP)

主要的核心思想：输入一组点云，然后随机选取最后保留下来的点云，通过加权距离去调整输出点云的位置

更新点云：

#ifndef DDX_WLOP_HEADER
#define DDX_WLOP_HEADER
#pragma once #include <vector>
#include <memory>
#include <Eigen/Dense>
#include "KDTreeFlann.h"
#include "KDTreeSearchParam.h"
#include <random>
#include <omp.h>
#include <iostream>
#include <fstream>namespace DDX {class WLOP
{
public:EIGEN_MAKE_ALIGNED_OPERATOR_NEW;
private:double mRadius;double mMu;int mSelect_percentage;bool require_uniform_sampling;std::vector<Eigen::Vector3d>& original_point;std::vector<Eigen::Vector3d> sample_points ;public:WLOP(std::vector<Eigen::Vector3d>& point , double radius ,int percentage  ,double mu = 0.45 ,bool density = false):original_point(point),mRadius (radius) ,mMu(mu) ,mSelect_percentage(percentage) ,require_uniform_sampling (density){};void Run (int iter_number) {size_t number_of_original = original_point.size() ;size_t number_of_sample = static_cast<size_t> (float(number_of_original) * (mSelect_percentage / float(100.0))) ;std::vector<Eigen::Vector3d> update_sample_points(number_of_sample);// 1 按照百分比生成下采样的点sample_points.reserve (number_of_sample);generate_sample_points(number_of_original ,number_of_sample);std::cout << "number_of_sample " << number_of_sample << std::endl;std::cout << "sample_points size " << sample_points.size() << std::endl;// 2 初始化一个原始的kd tree const auto original_kd_tree = create_Kdtree (original_point);std::vector<double> original_density (number_of_original);std::vector<double> samples_density (number_of_sample);if (require_uniform_sampling) {for (int i = 0 ; i < (int)original_point.size() ;++i){original_density[i] = compute_density_weight_for_original_point (original_kd_tree ,original_point[i] ,mRadius);}}// 3 迭代更新for (int iter_n = 0; iter_n < iter_number; ++iter_n) {std::cout << "iter " << iter_n << std::endl;// 创建采样点的kd tree const auto sample_kd_tree = create_Kdtree (sample_points);for (int i = 0 ; i < (int)sample_points.size() ; ++i){samples_density[i] = compute_density_weight_for_sample_point (sample_kd_tree ,sample_points[i] ,mRadius);}for (int i = 0 ; i < (int)sample_points.size() ; ++i) {update_sample_points[i] = compute_update_sample_point (i ,original_kd_tree ,sample_kd_tree ,original_density ,samples_density ,mRadius);}for (int i = 0 ; i < (int)update_sample_points.size() ; ++i) {sample_points[i] = update_sample_points[i];}}} ;void write_obj (const std::string& path) {std::ofstream delaunay3d (path.c_str());for (int i = 0 ; i < sample_points.size() ; ++i) {delaunay3d << "v " << sample_points[i][0] << " "<< sample_points[i][1] << " "<< sample_points[i][2] << std::endl ;}delaunay3d.close ();}
private:    std::unique_ptr<KDTreeFlann> create_Kdtree (std::vector<Eigen::Vector3d>& point) {Eigen::MatrixXd dataset0 = Eigen::Map<Eigen::MatrixXd>(reinterpret_cast<double*>(point.data()),3,point.size());return std::make_unique<KDTreeFlann> (dataset0) ;};void generate_sample_points (size_t number_of_original, size_t number_of_sample ) {size_t diff_number = number_of_original - number_of_sample;std::random_device rd ; std::mt19937 gen(rd()) ; constexpr int int_max = std::numeric_limits<int>::max () ;std::uniform_int_distribution<int> dis(int(0),int_max) ;std::vector<int> vist (number_of_original ,-1);int count_indx = 0;while (count_indx < diff_number ){int randint = dis (gen) ;const auto i13x = randint % (int)number_of_original ;if (vist[i13x] < 0) {vist[i13x] = 1;count_indx++;}}for (size_t i = 0 ; i < number_of_original ; ++i) {if (vist[i] < 0)sample_points.push_back (original_point[i]);}}double compute_density_weight_for_original_point (const std::unique_ptr<KDTreeFlann>& original_kd_tree ,Eigen::Vector3d const&p ,double radius ){double radius2 = radius * radius;double density_weight = (double)1.0;double iradius16 = -(double)4.0 / radius2;std::vector<int> index ;std::vector<double> distance ;const auto k = original_kd_tree->SearchRadius (p ,radius ,index ,distance) ;for (int i = 0 ; i < k ; ++i) {if (distance[i] < 1e-8) continue;density_weight += std::exp(distance[i] * iradius16);}return double(1.0) / density_weight;};double compute_density_weight_for_sample_point (const std::unique_ptr<KDTreeFlann>& sample_kd_tree  ,Eigen::Vector3d const&p ,double radius){//Compute density weightdouble radius2 = radius * radius;double density_weight = (double)1.0;double iradius16 = -(double)4.0 / radius2;std::vector<int> index ;std::vector<double> distance ;const auto k = sample_kd_tree->SearchRadius (p ,radius ,index ,distance) ;for (int i = 0 ; i < k ; ++i) {density_weight += std::exp(distance[i] * iradius16);}return density_weight;};Eigen::Vector3d compute_update_sample_point (int indx ,const std::unique_ptr<KDTreeFlann>& original_kd_tree,const std::unique_ptr<KDTreeFlann>& sample_kd_tree ,const std::vector<double>& sample_densities ,const std::vector<double> original_densities ,double radius) {bool is_original_densities_empty = original_densities.empty();bool is_sample_densities_empty = sample_densities.empty();double radius2 = radius * radius;double iradius16 = -(double)4.0 / radius2;// Compute Average TermEigen::Vector3d average (0 , 0 ,0);if (true) {    double average_weight_sum = (double)0.0;std::vector<int> index ;std::vector<double> distance ; const auto k = original_kd_tree->SearchRadius (sample_points[indx] , radius ,index ,distance);for (int j = 0 ; j < k ; j++){double dist2 = distance[j];if (dist2 < 1e-10) continue;double weight = (double)0.0;weight = std::exp(dist2 * iradius16);if (!is_original_densities_empty){weight *= original_densities[index[j]];}average_weight_sum += weight;average += original_point[index[j]] * weight;}if (k == 0 || average_weight_sum < 1e-20){average = sample_points[indx] ;}else{average = average / average_weight_sum;}}//Compute repulsion termEigen::Vector3d repulsion (0.0 , 0.0 ,0.0);if (true) {double weight = (double)1.0;double repulsion_weight_sum = (double)0.0;std::vector<int> index ;std::vector<double> distance ; const auto k = sample_kd_tree->SearchRadius (sample_points[indx] , radius ,index ,distance);for (int j = 0 ; j < k ; j++){double dist2 = distance[j];if (dist2 < 1e-10) continue;double dist = std::sqrt(dist2);weight = std::exp(dist2 * iradius16) * std::pow(double(1.0) / dist ,2); // L1if (!is_sample_densities_empty){weight *= sample_densities[index[j]];}Eigen::Vector3d diff = sample_points[indx] - sample_points[index[j]];repulsion_weight_sum += weight;repulsion = repulsion + diff * weight;}if (k < 3 || repulsion_weight_sum < double(1e-20)) {repulsion = Eigen::Vector3d(0.0 ,0.0 ,0.0);}else {repulsion = repulsion / repulsion_weight_sum;}}return average + mMu * repulsion ;  };};}#endif /* DDX_WLOP_HEADER */

githut: GitHub - xiandenis/WLOP: Consolidation of Unorganized Point Clouds for Surface Reconstruction

Consolidation of Unorganized Point Clouds for Surface Reconstruction(WLOP)的代码相关推荐

Learning Occupancy Function from Point Clouds for Surface Reconstruction论文笔记
Motivation&Contribution recent implicit function-based shape-learning methods focus only more on ...
泊松重建（Poisson Surface Reconstruction）
简介泊松重建是Michael Kazhdan等在2006年提出的网格重建方法,其文章题目为"Poisson Surface Reconstruction". Poisson-Re ...
Screened Poisson Surface Reconstruction (Version 8.0)
转自:https://www.cs.jhu.edu/~misha/Code/PoissonRecon/Version8.0/ linksexecutablesusagechanges LINKS Pa ...
[CVPR2021 Oral] PREDATOR: Registration of 3D Point Clouds with Low Overlap 损失函数含代码理解
PREDATOR: Registration of 3D Point Clouds with Low Overlap 损失函数含代码理解 Overlap loss,Matchability loss ...
Surface reconstruction from unorganized points
一．原论文链接本文描述了一种算法,输入未知实体M上或者附近的无序的数据点集合{x1,x2,-,xn},建立一个隐式函数代表点到实体M的符号距离,最后提取函数的零等值面,得到实体M的近似曲面. 1. ...
Poisson Surface Reconstruction
http://vr.tu-freiberg.de/scivi/?page_id=365 转载于:https://www.cnblogs.com/guochen/p/6567453.html
Screened Poisson Surface Reconstruction
http://www.cs.jhu.edu/~misha/Code/PoissonRecon/Version9.01/ 转载于:https://www.cnblogs.com/guochen/p/65 ...
图形学必备！斯坦福图形学几何处理算法课程1——Surface Reconstruction
表面重建,通过扫描设备扫描获得点云,之后重建下面就是具体的扫描系统,问题是要在两个图片上匹配特征激光脉冲,适用于远距离大物体, 依赖于测量时间的准确性目标是从扫描集重建物体方法一是以一个扫描覆 ...
经典点云去噪算法总结
以下列出方法均有运行成功的代码,所有工程文件我都会放在如下链接: https://github.com/gx-sun/classic-point-cloud-denoising-methods 欢迎s ...
图形处理（九）点云重建（下）法矢求取、有向距离场等值面提取
一.相关理论在点云重建算法中,点云法矢的求取是非常重要的一步.之前导师让我做点云尖锐特征边重建时,发了一堆异向法矢求取的英文paper,当时我很迷糊,就问了老师,让我做点云重建,为什么发的文献资料都 ...

Consolidation of Unorganized Point Clouds for Surface Reconstruction(WLOP)的代码

Consolidation of Unorganized Point Clouds for Surface Reconstruction(WLOP)的代码相关推荐

最新文章

热门文章