10

我有一个应用程序,它在图像/照片之上定义了一个真实世界的矩形,当然在 2D 中它可能不是一个矩形,因为你是从一个角度看它。

问题是,假设矩形需要在其上绘制网格线,例如,如果它是 3x5,那么我需要从第 1 面到第 3 面绘制 2 条线,从第 2 面到第 4 面绘制 4 条线。

截至目前,我将每条线分成等距的部分,以获得所有网格线的起点和终点。然而,矩形的角度越大,这些线就越“不正确”,因为离你越远的水平线应该更靠近。

有谁知道我应该搜索的算法的名称?

是的,我知道您可以在 3D 中执行此操作,但是对于此特定应用程序,我仅限于 2D。

4

9 回答 9

15

这是解决方案。

基本思想是您可以通过对角连接角找到矩形的正确透视“中心”。两条结果线的交点是您的透视正确中心。从那里您将矩形细分为四个较小的矩形,然后重复该过程。次数取决于您想要的准确度。您可以细分到像素大小以下,以获得有效的完美透视。

然后在您的子矩形中,您只需应用标准未校正的“纹理”三角形或矩形或其他任何东西。

您可以执行此算法,而无需费心构建“真实”3d 世界。如果您确实有一个真实的 3D 世界建模,但您的纹理三角形没有在硬件中进行透视校正,或者您需要一种高性能的方法来获得透视正确的平面,而无需每像素渲染技巧,这也很有用。

于 2009-02-10T04:02:48.343 回答
7

图片描述 图片:双线性和透视变换的示例(注意:上下水平网格线的高度实际上是其余线高度的一半,在两张图中)

=========================================

我知道这是一个老问题,但我有一个通用的解决方案,所以我决定发布它,希望它对未来的读者有用。下面的代码可以绘制任意透视网格,无需重复计算。

我实际上是从一个类似的问题开始的:绘制一个 2D 透视网格,然后转换下划线图像以恢复透视。

我开始在这里阅读: http ://www.imagemagick.org/Usage/distorts/#bilinear_forward

然后在这里(Leptonica 图书馆): http: //www.leptonica.com/affine.html

我发现了这个:

当您在有限距离内从某个任意方向查看平面上的物体时,您会在图像中获得额外的“梯形失真”失真。这是一种投影变换,它保持直线笔直,但不保留线之间的角度。这种扭曲不能用线性仿射变换来描述,事实上,分母中的 x 和 y 相关项不同。

正如许多人已经在此线程中指出的那样,这种转换不是线性的。它涉及求解一个由 8 个方程组成的线性系统(一次)以计算 8 个所需的系数,然后您可以使用它们来转换任意数量的点。

为了避免在我的项目中包含所有 Leptonica 库,我从中提取了一些代码,删除了所有特殊的 Leptonica 数据类型和宏,修复了一些内存泄漏并将其转换为 C++ 类(主要是出于封装原因)只做一件事:它将 (Qt) QPointF float (x,y) 坐标映射到相应的透视坐标。

如果你想让代码适应另一个 C++ 库,唯一需要重新定义/替换的是 QPointF 坐标类。

我希望一些未来的读者会发现它有用。下面的代码分为3部分:

A.如何使用genImageProjective C++类绘制2D透视Grid的示例

B. genImageProjective.h 文件

C. genImageProjective.cpp 文件

//============================================================
// C++ Code Example on how to use the 
//     genImageProjective class to draw a perspective 2D Grid
//============================================================

#include "genImageProjective.h"

// Input: 4 Perspective-Tranformed points:
//        perspPoints[0] = top-left
//        perspPoints[1] = top-right
//        perspPoints[2] = bottom-right
//        perspPoints[3] = bottom-left
void drawGrid(QPointF *perspPoints)
{
(...)
        // Setup a non-transformed area rectangle
        // I use a simple square rectangle here because in this case we are not interested in the source-rectangle,
        //  (we want to just draw a grid on the perspPoints[] area)
        //   but you can use any arbitrary rectangle to perform a real mapping to the perspPoints[] area
        QPointF topLeft = QPointF(0,0);
        QPointF topRight = QPointF(1000,0);
        QPointF bottomRight = QPointF(1000,1000);
        QPointF bottomLeft = QPointF(0,1000);
        float width = topRight.x() - topLeft.x();
        float height = bottomLeft.y() - topLeft.y();

        // Setup Projective trasform object
        genImageProjective imageProjective;
        imageProjective.sourceArea[0] = topLeft;
        imageProjective.sourceArea[1] = topRight;
        imageProjective.sourceArea[2] = bottomRight;
        imageProjective.sourceArea[3] = bottomLeft;
        imageProjective.destArea[0] = perspPoints[0];
        imageProjective.destArea[1] = perspPoints[1];
        imageProjective.destArea[2] = perspPoints[2];
        imageProjective.destArea[3] = perspPoints[3];
        // Compute projective transform coefficients
        if (imageProjective.computeCoeefficients() != 0)
            return; // This can actually fail if any 3 points of Source or Dest are colinear

        // Initialize Grid parameters (without transform)
        float gridFirstLine = 0.1f; // The normalized position of first Grid Line (0.0 to 1.0)
        float gridStep = 0.1f;      // The normalized Grd size (=distance between grid lines: 0.0 to 1.0)

        // Draw Horizonal Grid lines
        QPointF lineStart, lineEnd, tempPnt;
        for (float pos = gridFirstLine; pos <= 1.0f; pos += gridStep)
        {
            // Compute Grid Line Start
            tempPnt = QPointF(topLeft.x(), topLeft.y() + pos*width);
            imageProjective.mapSourceToDestPoint(tempPnt, lineStart);
            // Compute Grid Line End
            tempPnt = QPointF(topRight.x(), topLeft.y() + pos*width);
            imageProjective.mapSourceToDestPoint(tempPnt, lineEnd);

            // Draw Horizontal Line (use your prefered method to draw the line)
            (...)
        }
        // Draw Vertical Grid lines
        for (float pos = gridFirstLine; pos <= 1.0f; pos += gridStep)
        {
            // Compute Grid Line Start
            tempPnt = QPointF(topLeft.x() + pos*height, topLeft.y());
            imageProjective.mapSourceToDestPoint(tempPnt, lineStart);
            // Compute Grid Line End
            tempPnt = QPointF(topLeft.x() + pos*height, bottomLeft.y());
            imageProjective.mapSourceToDestPoint(tempPnt, lineEnd);

            // Draw Vertical Line (use your prefered method to draw the line)
            (...)
        }
(...)
}

==========================================



//========================================
//C++ Header File: genImageProjective.h
//========================================

#ifndef GENIMAGE_H
#define GENIMAGE_H

#include <QPointF>

// Class to transform an Image Point using Perspective transformation
class genImageProjective
{
public:
    genImageProjective();

    int computeCoeefficients(void);
    int mapSourceToDestPoint(QPointF& sourcePoint, QPointF& destPoint);

public:
    QPointF sourceArea[4]; // Source Image area limits (Rectangular)
    QPointF destArea[4];   // Destination Image area limits (Perspectivelly Transformed)

private:
    static int gaussjordan(float  **a, float  *b, int n);

    bool coefficientsComputed;
    float vc[8];           // Vector of Transform Coefficients
};

#endif // GENIMAGE_H
//========================================


//========================================
//C++ CPP File: genImageProjective.cpp
//========================================

#include <math.h>
#include "genImageProjective.h"

// ----------------------------------------------------
// class genImageProjective
// ----------------------------------------------------
genImageProjective::genImageProjective()
{
    sourceArea[0] = sourceArea[1] = sourceArea[2] = sourceArea[3] = QPointF(0,0);
    destArea[0] = destArea[1] = destArea[2] = destArea[3] = QPointF(0,0);
    coefficientsComputed = false;
}


// --------------------------------------------------------------
// Compute projective transform coeeeficients
// RetValue: 0: Success, !=0: Error
/*-------------------------------------------------------------*
 *                Projective coordinate transformation         *
 *-------------------------------------------------------------*/
/*!
 *  computeCoeefficients()
 *
 *      Input:  this->sourceArea[4]: (source 4 points; unprimed)
 *              this->destArea[4]:   (transformed 4 points; primed)
 *              this->vc  (computed vector of transform coefficients)
 *      Return: 0 if OK; <0 on error
 *
 *  We have a set of 8 equations, describing the projective
 *  transformation that takes 4 points (sourceArea) into 4 other
 *  points (destArea).  These equations are:
 *
 *          x1' = (c[0]*x1 + c[1]*y1 + c[2]) / (c[6]*x1 + c[7]*y1 + 1)
 *          y1' = (c[3]*x1 + c[4]*y1 + c[5]) / (c[6]*x1 + c[7]*y1 + 1)
 *          x2' = (c[0]*x2 + c[1]*y2 + c[2]) / (c[6]*x2 + c[7]*y2 + 1)
 *          y2' = (c[3]*x2 + c[4]*y2 + c[5]) / (c[6]*x2 + c[7]*y2 + 1)
 *          x3' = (c[0]*x3 + c[1]*y3 + c[2]) / (c[6]*x3 + c[7]*y3 + 1)
 *          y3' = (c[3]*x3 + c[4]*y3 + c[5]) / (c[6]*x3 + c[7]*y3 + 1)
 *          x4' = (c[0]*x4 + c[1]*y4 + c[2]) / (c[6]*x4 + c[7]*y4 + 1)
 *          y4' = (c[3]*x4 + c[4]*y4 + c[5]) / (c[6]*x4 + c[7]*y4 + 1)
 *
 *  Multiplying both sides of each eqn by the denominator, we get
 *
 *           AC = B
 *
 *  where B and C are column vectors
 *
 *         B = [ x1' y1' x2' y2' x3' y3' x4' y4' ]
 *         C = [ c[0] c[1] c[2] c[3] c[4] c[5] c[6] c[7] ]
 *
 *  and A is the 8x8 matrix
 *
 *             x1   y1     1     0   0    0   -x1*x1'  -y1*x1'
 *              0    0     0    x1   y1   1   -x1*y1'  -y1*y1'
 *             x2   y2     1     0   0    0   -x2*x2'  -y2*x2'
 *              0    0     0    x2   y2   1   -x2*y2'  -y2*y2'
 *             x3   y3     1     0   0    0   -x3*x3'  -y3*x3'
 *              0    0     0    x3   y3   1   -x3*y3'  -y3*y3'
 *             x4   y4     1     0   0    0   -x4*x4'  -y4*x4'
 *              0    0     0    x4   y4   1   -x4*y4'  -y4*y4'
 *
 *  These eight equations are solved here for the coefficients C.
 *
 *  These eight coefficients can then be used to find the mapping
 *  (x,y) --> (x',y'):
 *
 *           x' = (c[0]x + c[1]y + c[2]) / (c[6]x + c[7]y + 1)
 *           y' = (c[3]x + c[4]y + c[5]) / (c[6]x + c[7]y + 1)
 *
 */
int genImageProjective::computeCoeefficients(void)
{
    int retValue = 0;
    int     i;
    float  *a[8];  /* 8x8 matrix A  */
    float  *b = this->vc; /* rhs vector of primed coords X'; coeffs returned in vc[] */

    b[0] = destArea[0].x();
    b[1] = destArea[0].y();
    b[2] = destArea[1].x();
    b[3] = destArea[1].y();
    b[4] = destArea[2].x();
    b[5] = destArea[2].y();
    b[6] = destArea[3].x();
    b[7] = destArea[3].y();

    for (i = 0; i < 8; i++)
        a[i] = NULL;
    for (i = 0; i < 8; i++)
    {
        if ((a[i] = (float *)calloc(8, sizeof(float))) == NULL)
        {
            retValue = -100; // ERROR_INT("a[i] not made", procName, 1);
            goto Terminate;
        }
    }

    a[0][0] = sourceArea[0].x();
    a[0][1] = sourceArea[0].y();
    a[0][2] = 1.;
    a[0][6] = -sourceArea[0].x() * b[0];
    a[0][7] = -sourceArea[0].y() * b[0];
    a[1][3] = sourceArea[0].x();
    a[1][4] = sourceArea[0].y();
    a[1][5] = 1;
    a[1][6] = -sourceArea[0].x() * b[1];
    a[1][7] = -sourceArea[0].y() * b[1];
    a[2][0] = sourceArea[1].x();
    a[2][1] = sourceArea[1].y();
    a[2][2] = 1.;
    a[2][6] = -sourceArea[1].x() * b[2];
    a[2][7] = -sourceArea[1].y() * b[2];
    a[3][3] = sourceArea[1].x();
    a[3][4] = sourceArea[1].y();
    a[3][5] = 1;
    a[3][6] = -sourceArea[1].x() * b[3];
    a[3][7] = -sourceArea[1].y() * b[3];
    a[4][0] = sourceArea[2].x();
    a[4][1] = sourceArea[2].y();
    a[4][2] = 1.;
    a[4][6] = -sourceArea[2].x() * b[4];
    a[4][7] = -sourceArea[2].y() * b[4];
    a[5][3] = sourceArea[2].x();
    a[5][4] = sourceArea[2].y();
    a[5][5] = 1;
    a[5][6] = -sourceArea[2].x() * b[5];
    a[5][7] = -sourceArea[2].y() * b[5];
    a[6][0] = sourceArea[3].x();
    a[6][1] = sourceArea[3].y();
    a[6][2] = 1.;
    a[6][6] = -sourceArea[3].x() * b[6];
    a[6][7] = -sourceArea[3].y() * b[6];
    a[7][3] = sourceArea[3].x();
    a[7][4] = sourceArea[3].y();
    a[7][5] = 1;
    a[7][6] = -sourceArea[3].x() * b[7];
    a[7][7] = -sourceArea[3].y() * b[7];

    retValue = gaussjordan(a, b, 8);

Terminate:
    // Clean up
    for (i = 0; i < 8; i++)
    {
        if (a[i])
            free(a[i]);
    }

    this->coefficientsComputed = (retValue == 0);
    return retValue;
}


/*-------------------------------------------------------------*
 *               Gauss-jordan linear equation solver           *
 *-------------------------------------------------------------*/
/*
 *  gaussjordan()
 *
 *      Input:   a  (n x n matrix)
 *               b  (rhs column vector)
 *               n  (dimension)
 *      Return:  0 if ok, 1 on error
 *
 *      Note side effects:
 *            (1) the matrix a is transformed to its inverse
 *            (2) the vector b is transformed to the solution X to the
 *                linear equation AX = B
 *
 *      Adapted from "Numerical Recipes in C, Second Edition", 1992
 *      pp. 36-41 (gauss-jordan elimination)
 */
#define  SWAP(a,b)   {temp = (a); (a) = (b); (b) = temp;}
int genImageProjective::gaussjordan(float  **a, float  *b, int n)
{
    int retValue = 0;
    int i, icol=0, irow=0, j, k, l, ll;
    int *indexc = NULL, *indexr = NULL, *ipiv = NULL;
    float  big, dum, pivinv, temp;

    if (!a)
    {
        retValue = -1; // ERROR_INT("a not defined", procName, 1);
        goto Terminate;
    }
    if (!b)
    {
        retValue = -2; // ERROR_INT("b not defined", procName, 1);
        goto Terminate;
    }

    if ((indexc = (int *)calloc(n, sizeof(int))) == NULL)
    {
        retValue = -3; // ERROR_INT("indexc not made", procName, 1);
        goto Terminate;
    }
    if ((indexr = (int *)calloc(n, sizeof(int))) == NULL)
    {
        retValue = -4; // ERROR_INT("indexr not made", procName, 1);
        goto Terminate;
    }
    if ((ipiv = (int *)calloc(n, sizeof(int))) == NULL)
    {
        retValue = -5; // ERROR_INT("ipiv not made", procName, 1);
        goto Terminate;
    }

    for (i = 0; i < n; i++)
    {
        big = 0.0;
        for (j = 0; j < n; j++)
        {
            if (ipiv[j] != 1)
            {
                for (k = 0; k < n; k++)
                {
                    if (ipiv[k] == 0)
                    {
                        if (fabs(a[j][k]) >= big)
                        {
                            big = fabs(a[j][k]);
                            irow = j;
                            icol = k;
                        }
                    }
                    else if (ipiv[k] > 1)
                    {
                        retValue = -6; // ERROR_INT("singular matrix", procName, 1);
                        goto Terminate;
                    }
                }
            }
        }
        ++(ipiv[icol]);

        if (irow != icol)
        {
            for (l = 0; l < n; l++)
                SWAP(a[irow][l], a[icol][l]);
            SWAP(b[irow], b[icol]);
        }

        indexr[i] = irow;
        indexc[i] = icol;
        if (a[icol][icol] == 0.0)
        {
            retValue = -7; // ERROR_INT("singular matrix", procName, 1);
            goto Terminate;
        }
        pivinv = 1.0 / a[icol][icol];
        a[icol][icol] = 1.0;
        for (l = 0; l < n; l++)
            a[icol][l] *= pivinv;
        b[icol] *= pivinv;

        for (ll = 0; ll < n; ll++)
        {
            if (ll != icol)
            {
                dum = a[ll][icol];
                a[ll][icol] = 0.0;
                for (l = 0; l < n; l++)
                    a[ll][l] -= a[icol][l] * dum;
                b[ll] -= b[icol] * dum;
            }
        }
    }

    for (l = n - 1; l >= 0; l--)
    {
        if (indexr[l] != indexc[l])
        {
            for (k = 0; k < n; k++)
                SWAP(a[k][indexr[l]], a[k][indexc[l]]);
        }
    }

Terminate:
    if (indexr)
        free(indexr);
    if (indexc)
        free(indexc);
    if (ipiv)
        free(ipiv);
    return retValue;
}


// --------------------------------------------------------------
// Map a source point to destination using projective transform
// --------------------------------------------------------------
// Params:
//  sourcePoint: initial point
//  destPoint:   transformed point
// RetValue: 0: Success, !=0: Error
// --------------------------------------------------------------
//  Notes:
//   1. You must call once computeCoeefficients() to compute
//      the this->vc[] vector of 8 coefficients, before you call
//      mapSourceToDestPoint().
//   2. If there was an error or the 8 coefficients were not computed,
//      a -1 is returned and destPoint is just set to sourcePoint value.
// --------------------------------------------------------------
int genImageProjective::mapSourceToDestPoint(QPointF& sourcePoint, QPointF& destPoint)
{
    if (coefficientsComputed)
    {
        float factor = 1.0f / (vc[6] * sourcePoint.x() + vc[7] * sourcePoint.y() + 1.);
        destPoint.setX( factor * (vc[0] * sourcePoint.x() + vc[1] * sourcePoint.y() + vc[2]) );
        destPoint.setY( factor * (vc[3] * sourcePoint.x() + vc[4] * sourcePoint.y() + vc[5]) );
        return 0;
    }
    else // There was an error while computing coefficients
    {
        destPoint = sourcePoint; // just copy the source to destination...
        return -1;               // ...and return an error
    }
}
//========================================
于 2012-11-22T07:34:25.363 回答
3

使用 Breton 的细分方法(与 Mongo 的扩展方法有关),将为您提供准确的任意二次幂除法。要使用这些方法拆分为非二次幂分割,您必须细分为子像素间距,这在计算上可能很昂贵。

但是,我相信您可以将Haga 定理的变体(在折纸中用于将一边划分为 (N-1) 个的边分为 N 个)到透视正方形细分以产生任意划分最接近 2 的幂,而无需继续细分。

于 2009-02-10T04:27:20.390 回答
2

最优雅和最快的解决方案是找到将矩形坐标映射到照片坐标的单应矩阵。

有了一个像样的矩阵库,只要你知道你的数学,它应该不是一项艰巨的任务。

关键词:准线,单应性,直接线性变换

然而,上面的递归算法应该可以工作,但如果你的资源有限,投影几何是唯一的方法。

于 2012-01-31T06:19:57.190 回答
1

我认为所选答案不是可用的最佳解决方案。一个更好的解决方案是将矩形的透视(投影)变换应用于简单的网格,如下所示 Matlab 脚本和图像显示。您也可以使用 C++ 和 OpenCV 实现此算法。

function drawpersgrid
sz      = [ 24, 16 ]; % [x y]
srcpt   = [ 0 0; sz(1) 0; 0 sz(2); sz(1) sz(2)];
destpt  = [ 20 50; 100 60; 0 150; 200 200;];

% make rectangular grid
[X,Y]   = meshgrid(0:sz(1),0:sz(2));

% find projective transform matching corner points
tform   = maketform('projective',srcpt,destpt);

% apply the projective transform to the grid
[X1,Y1] = tformfwd(tform,X,Y);

hold on;

%% find grid

for i=1:sz(2)
    for j=1:sz(1)
        x = [ X1(i,j);X1(i,j+1);X1(i+1,j+1);X1(i+1,j);X1(i,j)];
        y = [ Y1(i,j);Y1(i,j+1);Y1(i+1,j+1);Y1(i+1,j);Y1(i,j)];
        plot(x,y,'b');
    end
end
hold off;

投影网格

于 2017-05-13T19:06:21.890 回答
0

在特殊情况下,当您垂直于边 1 和 3 时,您可以将这些边分成相等的部分。然后画一条对角线,通过对角线的每个交点和前面画的分界线画出与边 1 的平行线。

于 2009-02-13T13:23:00.100 回答
0

这是我想出来的几何解决方案。我不知道“算法”是否有名字。

假设您想首先将“矩形”划分为 n 个带有垂直线的部分。

目标是将点 P1..Pn-1 放在顶线上,我们可以用它来画线穿过它们到左右线相交或平行的点,当这样的点不存在时。

如果顶线和底线彼此平行,只需放置这些点以等距分割角之间的顶线。

否则将 n 个点 Q1..Qn 放在左侧线上,使这些点和左上角等距,并且 i < j => Qi 比 Qj 更靠近左上角。为了将 Q 点映射到顶线,找到从 Qn 到右上角的线的交点 S,并通过顶线和底线的交点找到与左线平行的线。现在将 S 与 Q1..Qn-1 连接起来。新线与顶线的交点是所需的 P 点。

对水平线做这个模拟。

于 2014-04-14T15:51:42.700 回答
0

给定绕 y 轴的旋转,特别是如果旋转表面是平面的,则透视是由垂直梯度生成的。这些在视角上逐渐接近。而不是使用对角线来定义四个矩形,它可以在给定二的幂的情况下工作......定义两个矩形,左右。如果继续将表面划分为更窄的垂直部分,它们最终会高于宽度。这可以适应非方形的表面。如果围绕 x 轴旋转,则需要水平渐变。

于 2015-09-30T07:26:48.343 回答
-1

您需要做的是在 3D(世界)中表示它,然后将其投影到 2D(屏幕)。

这将要求您使用 4D 变换矩阵,该矩阵将 4D 齐次投影到 3D 齐次向量,然后您可以将其向下转换为 2D 屏幕空间向量。

我在谷歌也找不到它,但是一本好的计算机绘图书会有详细信息。

关键词是投影矩阵、投影变换、仿射变换、齐次向量、世界空间、屏幕空间、透视变换、3D变换

顺便说一句,这通常需要几个讲座来解释所有这些。祝你好运。

于 2009-02-09T23:31:07.920 回答