Question

I am porting some code that is using OpenCV just to fit an ellipse. I may rewrite it later but for now I need to use the OpenCV routine.
This code is just a few lines but I had to change almost everything because the profile of cvFitEllipse2() has changed.

My problematic is the following one: convert an array of point (pair of uchar) into something I can feed cvFitEllipse2() with

I did not found a lot of code sample and I find that OpenCV is poorly documented (but since I am new to OpenCV, maybe this feeling is wrong).
Anyway here is what I wrote:

/* centers is the input uchar array */
CvBox2D box;    //output
CvMemStorage* storage = cvCreateMemStorage(0);
CvSeq* seq = cvCreateSeq(CV_32FC2, sizeof(CvSeq), sizeof(CvPoint2D32f), storage);

for(int h=0; h<mm; ++h) 
{
    CvPoint2D32f p;
    p.x = (float)centers[h].x;
    p.y = (float)centers[h].y;
    cvSeqPush(seq, &p);
}

// Fits ellipse to current contour.
box = cvFitEllipse2(seq);

Problem: this code is damn slow compared to the original one.

• Is there any way to preallocate the memory required for seq? (using
  the CvMemStorage object)
• Are there better (simpler? containers than
  CvSeq?
• Can I replace the cvSeqPush by something faster?

Thanks a lot for your help!

Answer 1

I assume centers is a vector of points.

If you are happy to use the C++ interface instead of the old one, your pain is over.

Depending on the way centers is declared you can:

vector<cv::Point> centers;
RotatedRect rect = fitEllipse(centers);

Or the hard-core approach:

struct {
    uchar x, y
} myPoint;

vector<myPoint> centers;
...
int sz = centers.size();

// if you have a different type for x and y, modify here to CV_MY_TYPE
// Here I have taken a pointer to the vector data storage, 
// and fed it into the Mat.
// Check Mat documentation on constructors
Mat pts(1,sz,CV_8UC2, (void*)&centers[0].x); 
RotatedRect rect = fitEllipse(pts);
...
// take care not to change/ delete your vector 
// before finishing with the pts Mat.

EDIT

I have modified types to UCHAR