GreedyProjectionTriangulation class resulting grid normals are reversed

When I used this class to generate grids, I found that some grids were found to be opposite, resulting in the display effect of exported STL file being different from that of PCL visual window. I would like to ask if there is any method to repair the normal line

And I found that when using the same method of normal estimation, neither Poisson nor moving cube will have the opposite situation of normal, only greedy projection method will

PCL visualization

STL file display

` std::cout << "Using mls method estimation..." << endl;;
pcl::PointCloudpcl::PointNormal mls_points1;
pcl::PointCloudpcl::PointNormal::Ptr mls_points(new pcl::PointCloudpcl::PointNormal());
pcl::MovingLeastSquares<pcl::PointXYZ, pcl::PointNormal> mls;
cout << "MLS前点云数:" << cloudTranslated->points.size() << endl;
mls.setComputeNormals(true);
mls.setInputCloud(cloudTranslated);

    mls.setPolynomialOrder(true);
    // mls.setDilationIterations(10);
     //mls.setDilationVoxelSize(0.5);
     //mls.setSqrGaussParam(2.0);
     //mls.setUpsamplingRadius(5);
    mls.setPolynomialOrder (3); 
     //mls.setPointDensity(30);
    mls.setSearchMethod(kdtree_for_points);
    mls.setSearchRadius(3);
    mls.process(*mls_points);

    pcl::PointCloud<pcl::PointXYZ>::Ptr temp(new pcl::PointCloud<pcl::PointXYZ>());

    for (int i = 0; i < mls_points->points.size(); i++) {

        pcl::PointXYZ pt;
        pt.x = cloud->points[i].x;
        pt.y = cloud->points[i].y;
        pt.z = cloud->points[i].z;

        temp->points.push_back(pt);
    }

    cout << "MLS后点云数目 ：" << mls_points->points.size() << endl;
    cout << "temp ：" << temp->points.size() << endl;
    pcl::concatenateFields(*temp, *mls_points, *cloud_with_normals);
    std::cout << "移动最小二乘法线估计完成" << std::endl;
    cout << "最小二乘cloud_with_normals：" << cloud_with_normals->points.size() << endl;
   
    pcl::GreedyProjectionTriangulation<pcl::PointNormal> gp3;
    gp3.setSearchRadius(search_radius);//It was 0.025
    gp3.setMu(setMU); //It was 2.5
    gp3.setMaximumNearestNeighbors(maxiNearestNeighbors);    //It was 100
    gp3.setMaximumSurfaceAngle(M_PI/2); // 45 degrees    //it was 4
    gp3.setMinimumAngle(M_PI/4); // 10 degrees //It was 18
    gp3.setMaximumAngle(M_PI/1.2); // 120 degrees        //it was 1.5
    gp3.setNormalConsistency(false); //It was false
    gp3.setConsistentVertexOrdering(true);
    gp3.setInputCloud(cloud_with_normals);
    gp3.setSearchMethod(kdtree_for_normals);
    gp3.reconstruct(triangles);
std::string output_dir1 = "D://Program Files//CloudPoint//model_converted_pcd//dragon_mesh.stl";
		pcl::io::savePolygonFileSTL(output_dir1.c_str(), triangles, true);

`

Start with the GreedyProjectionTriangulation configuration and its reconstruct call, using the provided MovingLeastSquares normal-estimation pipeline as the reproduction. Compare the generated triangles and normals with the PCL visualization and exported STL; done means the cause is identified and the resulting grid orientation is consistent between both views.