pthomet's profile - activity

Note 1 : this answer was edited following Eduardo's edit of his own answer. The initial unstable results under OSX were due to an error in the code. Sorry about this.

Note 2 : I made a more thorough analysis, that compares several strategies (after having added the distorsion coeffs of the camera). The code would not fit here, but it can be found at https://github.com/pthom/TestSolvePnp

enum SolvePnpStrategy
{
  Strategy_MySolvePnp_Epnp, //an home baked adapter of epnp using the epnp library source code
  Strategy_MySolvePnpPosit, //an home baked adapter of cvPOSIT (deprecated "OpenCV 1" pose estimation method)
  Strategy_solvePnp_P3p,    // opencv SOLVEPNP_P3P method
  Strategy_solvePnp_Iterative_InitialGuess, // opencv SOLVEPNP_ITERATIVE method with an initial guess
  Strategy_solvePnp_Epnp //opencv SOLVEPNP_EPNP method
};

Based on my experimentations, the order of the 3d points and image points does matter. It has to be adapted depending upon the strategy !

• With opencv's SOLVEPNP_EPNP the error can go down to 23.03 pixel. The order of the points does matter
• With MySolvePnpEpnp (an home baked adapter of epnp using the epnp library source code), the error is about 6.742 pixels, and the order is important It is strange that this "rewrite" gives different results
• With MySolvePnpPosit, the error is about 4.911 pixels and the order is important (in other cases the reprojection error is about 1278 pixels !)
• With solvePnp_P3p (cv::SOLVEPNP_P3P) the error is about 0.02961 pixels and the order does not matter much
• With solvePnp_Iterative_InitialGuess (cv::SOLVEPNP_ITERATIVE) the error can be 0 pixels if a good initial extrinsic guess is given (otherwise don't hope for any convergence). The order does not matter much with SOLVEPNP_ITERATIVE.

Original answer below

Many thanks Eduardo for your thorough analysis, and the time you invested for it :-)

This is not an answer, but a recap of my results following your suggestions.

Note : my tests were run under OSX El Capitan. I did not test linux or windows yet. I have the same results as you :

• With cv::SOLVEPNP_DLS, cv::SOLVEPNP_EPNP, cv::SOLVEPNP_UPNP , the reprojection error is 44.8698 pixels
• With cv::SOLVEPNP_ITERATIVE the reprojection error is 8.94 pixels (except if you set useExtrinsicGuess to true and use null vectors as inital rotation and translation}, and the returned pose is behind the camera
• With cv::SOLVEPNP_P3P the reprojection error is about 0.21 pixels

I was much less sucessful than you : I saw unstable results when using the optimized version, and no success when disabling the optimizations.

• Using cv::SOLVEPNP_UPNP

  • optimized version : if you run the program several times, the result you get can be 4535.94, 711.716, 302630, nan or 0 !!!
  • unoptimized version : result is always 711.832

• Using cv::SOLVEPNP_DLS

  • optimized version: result vary. I obtained ...

Hi,

I have found a set of parameters for which solvePnP does not give correct results although :

• the image points are perfect coordinates obtained with projectPoints applied to the objectPoints (no error at all in the image points)
• the objects points consist of 4 non coplanar points
• the parameters are quite innocent looking
• cvPOSIT is perfectly able to reconstruct the pose, when solvePnp seems not to be able to do so

This problem occurs with opencv 2.4.1 and 3.1.0, under osx and windows

When I reproject the point using the rotation and translation provided by solvePnP, I have a total error of about 8 pixels (I have tried all flags : ITERATIVE, P3P and EPNP)

If instead I use the rotation and translation provided by cvPOSIT the reprojection error can be reduced to 0.00014 pixels !

Here is a short program that demonstrates the error : I simply instantiate a camera matrix, some object points and their corresponding image points obtained through projectPoints. Then I compare the reprojection and I get 8 pixels of error!

#include <opencv2/core/core.hpp>
#include <opencv2/calib3d/calib3d.hpp>
#include <iostream>

void TestSolvePnp()
{
  std::vector<cv::Point3f> objectPoints;

  objectPoints.push_back(cv::Point3f(0.f, 0.f, 0.f));
  objectPoints.push_back(cv::Point3f(62.3174629f, 15.7940502f, 0.819983721f));
  objectPoints.push_back(cv::Point3f(-62.1225319f, 15.7540569f, 0.819464564f));
  objectPoints.push_back(cv::Point3f(-0.372639507f, 16.4230633f,  -36.5060043f));

  cv::Mat_<double> cameraIntrinsicParams(cv::Size(3, 3));
  cameraIntrinsicParams = 0.;
  cameraIntrinsicParams(0, 0) = 3844.4400000000001f;
  cameraIntrinsicParams(1, 1) = 3841.0599999999999f;
  cameraIntrinsicParams(0, 2) = 640.f;
  cameraIntrinsicParams(1, 2) = 380.f;
  cameraIntrinsicParams(2, 2) = 1.f;

  cv::Mat_<double> distCoeffs(cv::Size(1, 4));
  distCoeffs = 0.f;


  cv::Mat_<double> rotation(cv::Size(3, 1));
  rotation(0,0) = 0.07015543380659847f;
  rotation(0,1) = 0.06922079477774973f;
  rotation(0,2) = -0.00254676088325f;

  cv::Mat_<double> translation(cv::Size(3, 1));
  translation(0,0) = -35.3236f;
  translation(0,1) = -48.1699f;
  translation(0,2) = 769.068f;

  std::vector<cv::Point2f> imagePoints;
  cv::projectPoints(objectPoints, rotation, translation, cameraIntrinsicParams, distCoeffs, imagePoints);


  cv::Mat_<double> rotation2(cv::Size(3, 1));
  cv::Mat_<double> translation2(cv::Size(3, 1));

  cv::solvePnP(objectPoints, imagePoints,                 
               cameraIntrinsicParams, distCoeffs,
               rotation2, translation2,
               //false,//useExtrinsicGuess

               //Choose solvepnp flag below
               //cv::SOLVEPNP_UPNP
               //cv::SOLVEPNP_DLS
               //cv::SOLVEPNP_EPNP
               //cv::SOLVEPNP_P3P
               cv::SOLVEPNP_ITERATIVE
  );

  std::vector<cv::Point2f> imagePoints_Reproj(3);
  cv::projectPoints(objectPoints, rotation2, translation2, cameraIntrinsicParams, distCoeffs, imagePoints_Reproj);

  float sum = 0.;
  for (size_t i = 0; i < imagePoints.size(); i++)
    sum += DistPoint(imagePoints_Reproj[i], imagePoints[i]);

  std::cout << "sum=" << sum << std::endl;
}

int main(int argc, char **argv)
{
  TestSolvePnp();
}

However, if I replace the call to solvePnp by a call to cvPOSIT, the reprojection error goes down to 0.00014 pixels !

Below is the implementation of a function named MySolvePnpPosit() that reproduces the behavior of solvePnP, but uses cvPOSIT internally => when I use it, the error goes down almost zero.

namespace
{
  void Posit_IdealCameraCoords(const std::vector<cv::Point3f> & objectPoints, const std::vector<cv::Point2f> & imagePoints,
             cv::Mat &outRotationEstimated, cv::Mat & outTranslationEstimated)
  {

    CvPoint2D32f * imagePoints_c = (CvPoint2D32f ...