ydsf16 / pnp_solver Goto Github PK

View Code? Open in Web Editor NEW

94.0 94.0 30.0 18 KB

Personal implementations of solvers for PnP problem, including DLT and EPnP.

Home Page: https://zhuanlan.zhihu.com/p/59070440

CMake 1.01% C++ 98.99%

pnp_solver's People

Contributors

Stargazers

Watchers

pnp_solver's Issues

关于使用Gauss-Newton求解过程中，雅克比矩阵J的构造

PnP_Solver/pnp_solver.cpp

Lines 530 to 534 in 835e358

 // [B11 B12 B22 B13 B23 B33 B14 B24 B34 B44] 

 J ( i, 0 ) = 2 * betas[0] * L ( i, 0 ) + betas[1]* L ( i, 1 ) + betas[2]*L ( i, 3 ) + betas[3]*L ( i, 6 ); 

 J ( i, 1 ) = betas[0] * L ( i, 1 ) + 2 * betas[1]* L ( i, 2 ) + betas[2]*L ( i, 3 ) + betas[3]*L ( i, 7 ); 

 J ( i, 2 ) = betas[0] * L ( i, 3 ) + betas[1]* L ( i, 4 ) + 2 * betas[2]*L ( i, 5 ) + betas[3]*L ( i, 8 ); 

 J ( i, 3 ) = betas[0] * L ( i, 6 ) + betas[1]* L ( i, 7 ) + betas[2]*L ( i, 8 ) + 2 * betas[3]*L ( i, 9 );

作者您好，我从您的知乎[PnP]PnP问题之EPnP解法这篇文章过来的，关于使用Gauss-Newton求解文章中公式19，您在文章中构造的雅克比矩阵J的形式如下：

但是似乎在您给出的代码中，关于J矩阵中的各个系数似乎并不全为2，似乎是下面这种形式：

哪一种写法是正确的呢？

EPnP: sign of the betas for N>1

Thanks for open sourcing this great repo :)

If you have the time, I am curious about one aspect of the EPnP implementation:
Why the sign of the betas when $N > 1$ should be decided as the way they are implemented?

E.g. when $N=2$ :

PnP_Solver/pnp_solver.cpp

Lines 436 to 446 in 835e358

 if ( b3[0] < 0 ) { 

 betas[0] = sqrt ( -b3[0] ); 

 betas[1] = ( b3[2] < 0 ) ? sqrt ( -b3[2] ) : 0.0; 

 } else { 

 betas[0] = sqrt ( b3[0] ); 

 betas[1] = ( b3[2] > 0 ) ? sqrt ( b3[2] ) : 0.0; 

 } 

 if ( b3[1] < 0 ) { 

 betas[0] = -betas[0]; 

 }

However, by definition, $\beta_{11} = \beta_1 * \beta_1 > 0$. Thereby, why an estimate $\beta_{11}<0$ is still considered valid (keeping it as an absolute value), and is not ignored (thus assuming that solutions is spurious)?

Thanks in advance.

Assertion error (bug)

Hi, I tried your code, its perfect! It worked well on my laptop except one place.
On line 278 pop_solver.cpp
C.block ( 3, 0, 1, 4 ) = Eigen::Vector4d ( 1.0, 1.0, 1.0, 1.0 );
When I run the program, it got stuck here. assertion error. rows and cols != problem.
my fix is C.block ( 3, 0, 1, 4 ) = Eigen::Vector4d ( 1.0, 1.0, 1.0, 1.0 ).transpose();

Good job! I found your result is much much better than OpenCV's epnp!

different result with Eigen 3.2.0 and Eigen 3.3.4

I ran a simple program in two systems. Mac with Eigen3.3.4 and Linux with Eigen3.2.0. I got super different results.

Here's my code.

Eigen::Matrix3d m_K = Eigen::Matrix3d::Identity();
    m_K(0, 0) = 672.062;
    m_K(1, 1) = 674.562;
    m_K(0, 2) = 398.838;
    m_K(1, 2) = 279.38;

    std::vector< Eigen::Vector3d > vecPts3d;
    std::vector< Eigen::Vector2d > vecPts2d;

    Eigen::Vector3d p3d0(492.681, 1.73013, 159.257);
    Eigen::Vector3d p3d1(493.724, -4.69578, 69.9289);
    Eigen::Vector3d p3d2(597.975, -10.3263, 9.82591);
    Eigen::Vector3d p3d3(623.276, -9.67869, -303.368);
    Eigen::Vector3d p3d4(502.611, -4.8012, 66.27);

    Eigen::Vector2d p2d0(507.6, 327.6);
    Eigen::Vector2d p2d1(624.248, 239.454);
    Eigen::Vector2d p2d2(222.731, 263.926);
    Eigen::Vector2d p2d3(432.868, 275.006);
    Eigen::Vector2d p2d4(561.6, 244.384);

    vecPts3d.emplace_back(p3d0);
    vecPts3d.emplace_back(p3d1);
    vecPts3d.emplace_back(p3d2);
    vecPts3d.emplace_back(p3d3);
    vecPts3d.emplace_back(p3d4);
    vecPts2d.emplace_back(p2d0);
    vecPts2d.emplace_back(p2d1);
    vecPts2d.emplace_back(p2d2);
    vecPts2d.emplace_back(p2d3);
    vecPts2d.emplace_back(p2d4);

    Eigen::Matrix3d R;
    Eigen::Vector3d t;

    solvePnPbyEPnP(m_K, vecPts3d, vecPts2d, R, t);
    std::cout << R << std::endl;
    std::cout << t.transpose() << std::endl;

With Mac&Eigen3.3.4, I got

-0.947487  0.0657636  -0.312958
 0.0680975   0.997673 0.00347973
  0.312459 -0.0180147   -0.94976
  521.61 -34.9744  13.0391

With Linux&Eigen3.2.0 I got

 -0.931891   0.142008  -0.333785
 -0.142497  -0.989524 -0.0231554
 -0.333577   0.025985   0.942365
-1.94579e+156 -6.07921e+156  2.88192e+158

Also I tried Mac&Eigen3.2.0, another differnt result.

-0.946419 0.0664846  0.316023
0.0906903  0.993916 0.0624984
-0.309945 0.0878099 -0.946691
 5336.94 -628.954  38868.5

ydsf16 / pnp_solver Goto Github PK

pnp_solver's People

Contributors

Stargazers

Watchers

Forkers

pnp_solver's Issues

关于使用Gauss-Newton求解过程中，雅克比矩阵J的构造

EPnP: sign of the betas for N>1

Assertion error (bug)

different result with Eigen 3.2.0 and Eigen 3.3.4

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

Jobs

	// [B11 B12 B22 B13 B23 B33 B14 B24 B34 B44]
	J ( i, 0 ) = 2 * betas[0] * L ( i, 0 ) + betas[1]* L ( i, 1 ) + betas[2]L ( i, 3 ) + betas[3]L ( i, 6 );
	J ( i, 1 ) = betas[0] * L ( i, 1 ) + 2 * betas[1]* L ( i, 2 ) + betas[2]L ( i, 3 ) + betas[3]L ( i, 7 );
	J ( i, 2 ) = betas[0] * L ( i, 3 ) + betas[1]* L ( i, 4 ) + 2 * betas[2]L ( i, 5 ) + betas[3]L ( i, 8 );
	J ( i, 3 ) = betas[0] * L ( i, 6 ) + betas[1]* L ( i, 7 ) + betas[2]L ( i, 8 ) + 2 betas[3]*L ( i, 9 );

	if ( b3[0] < 0 ) {
	betas[0] = sqrt ( -b3[0] );
	betas[1] = ( b3[2] < 0 ) ? sqrt ( -b3[2] ) : 0.0;
	} else {
	betas[0] = sqrt ( b3[0] );
	betas[1] = ( b3[2] > 0 ) ? sqrt ( b3[2] ) : 0.0;
	}

	if ( b3[1] < 0 ) {
	betas[0] = -betas[0];
	}