fitEllipse (Bugfix #3165)

The parameter min_eps seems to be choosen too large for some long contours. In some cases the variable "rp²" is quite small. Thus the check for condition "if( rp² > min_eps )" prevents the code line "rp² = sqrt(2.0 / rp²);" from being executed. This results in e very small value for the box width at the end of the fitting algorithm. Also the angle is computet wrongly.

Example:
std::vector<cv::Point> points;

points.push_back(cv::Point(618,   522));
    points.push_back(cv::Point(102,   551));
    points.push_back(cv::Point(1,  680));
    points.push_back(cv::Point(351,  680));
    points.push_back(cv::Point(920,  671));
    points.push_back(cv::Point(1911,  664));
    points.push_back(cv::Point(2421,  603));
    points.push_back(cv::Point(1902,  571));
    points.push_back(cv::Point(1116,  549));
    points.push_back(cv::Point(897,   552));

cv::RotatedRect box = cv::fitEllipse(cv::Mat(points));
    std::cout << "Box has width: " << box.size.width << " height: " << box.size.height << " angle: " << box.angle << std::endl;

Output of this code:
Box has width: 1.37751e-006 height: 145.668 angle: 0

Proposed solution:
Reduce the value of the constant min_eps. Currently it shows up like this: "const double min_eps = 1e-6;" Reduce ist to "const double min_eps = 1e-8;" if there are no further implications to the fitting algorithm.

Regards,
Torsten