OpenCV comes with a function to help you find subpixel corners. It uses the dot product technique to refine corners detected by other techniques, like the Shi-Tomasi corner detector. The function works iteratively, refining the corners till a termination criteria is reached.

Refining corners to subpixel level

The function that lets you calculate better corner positions is cvFindCornerSubPix:

void cvFindCornerSubPix(const CvArr* image,
                        CvPoint2D32f* corners,
                        int count,
                        CvSize win,
                        CvSize zero_zone,
                        CvTermCriteria criteria);

Before calling this function, you must use cvGoodFeaturesToTrack to find the approximate location of corners in an image. This function will then refine those estimates.

With that in mind, I'll go through each one here.

image - Quite basic, you pass the image you want to work on. This should be the same image you used in cvGoodFeaturesToTrack (or you'll end up with weird results)

corners - This is the array that holds the approximate corners initially. The function modifies this array with the refined corner positions.

count - Quite basic, the number of points in the above array. Same as cvGoodFeaturesToTrack

win_ - The technique used by this function requires several equations. This is done by using several pixels around the corner. win lets you set the size of the window from which these pixels are taken. Example: cvSize(5, 5)

zero_zone - Again, the technique used by this function solves several equations. The "solving" part is done using a matrix. This matrix is inverted to get a solution. However, some matrices are non-invertible. To prevent this, some pixels around the corner are ignored. zero_zone is that area.

In the above picture, the red pixel is the (integer based) corner. win has been set to 7x7. zero_zone has been set to 3x3. So, only the green pixels are used to generating new equations. The grey pixels are ignored.

criteria - This is used in several iterative algorithms of OpenCV. It lets you specify the type (CV_TERMCRIT_ITER or CV_TERMCRIT_EPS or both), the number of iterations and the desired accuracy.

Summary

OpenCV lets you easily refine integer based corners. A simple to use function that does all the heavy duty work of iteratively increasing the accuracy of the supplied corners.

Also, I've attached a sample project to demonstrate the function.