c++opencvcomputer-visionprojective-geometry

OpenCV findHomography giving inacurate result

I am having trouble with the findHomography function in OpenCV, it is not translating even my corner points correctly.

Here is the code I have to calculate the homography.

std::vector<cv::Point> srcPts = board.getCornerPoints();
// Image size : 640x480
// [158, 110;
//  488, 114;
//  507, 296;
//  100, 284]

const std::vector<cv::Point2f> destPts { 
        cv::Point2f(-0.5, 0.5),
        cv::Point2f(0.5, 0.5),
        cv::Point2f(0.5, -0.5),
        cv::Point2f(-0.5, -0.5),
};
cv::Mat H = cv::findHomography(srcPts, destPts);

The goal is to transform the points in the corners of the selected object to a plane between -0.5 and 0.5. However, when trying on the upper left corner with this code :

cv::Point pt = boardDetector.getCornerPoints()[0];
cv::Mat test = (cv::Mat_<double>(3, 1) << pt.x, pt.y, 1);
cv::Mat H_pt = H * test;

Instead of getting (-0.5, 0.5) I'm getting (-0.59675, 0.59675).

When changing the coordinates of the corner points to camera coordinates by multiplying them with the camera matrix the results are even worse too.

I have tried specifying the method used by findHomography myself but it didn't change anything.

Any ideas on why the result is inaccurate?

Thanks

Solution

  • After multiplication, you still need to project your 2D point onto the w=1 plane (vector being (x,y,w), w being the last coordinate). You do that by dividing the vector by that value in the vector.

    Or just use cv::perspectiveTransform(points, mat).