OpenCV: HoughCircles returns invalid circle parameters

Question

I'm using "OpenCV Version: 3.4.0" with Python 3.6 and the Spyder IDE (Spyder 3.2.4) on Anaconda Navigator.

When I call HoughCircles() with these parameters, it returns invalid circle data:

=====================================================================
circles = cv2.HoughCircles(gray, cv2.HOUGH_GRADIENT, 1.8, 30, 10, 310,minRadius = 35, maxRadius = 255)

len(circles): 4
circles: [[10.][ 0.][ 0.][ 0.]]
=====================================================================

When I call HoughCircles() with these parameters, it returns valid circle data:

=====================================================================
circles = cv2.HoughCircles(gray, cv2.HOUGH_GRADIENT, 1.8, 30, 10, 300,minRadius = 35, maxRadius = 255)

len(circles): 1
circles: [[[319.5 571.5 251. ]]]
=====================================================================

Why does HoughCircles() returns invalid data under some circumstances? Shouldn't the function return some type of error code, or maybe "None"?

Answer 1

I ran into something similar in C++ where HoughCircles() would return a list with a single [0,0,0] in it for certain images. The root cause was a bug in the RemoveOverlaps() method in OpenCV. Here is the bug report that I wrote:
https://github.com/opencv/opencv/issues/14432

This same issue also appears to affect Python, as described in this question:
OpenCV HoughCircles occasionally returning [0. 0. 0.]

You might try manually building a version of OpenCV with the patch applied to see if that helps (or a future release like 3.4.7 or 4.2.0). However, since you are getting [[10][0][0][0]] instead of [[[0 0 0]]], I think you are probably running into a different bug than I was. If so, I would suggest filing a new bug report about it.

Finally, whenever posting about computer vision, it is a good idea to include the image file that you are testing with, as it will be impossible to reproduce the issue without it.