I solved this in OpenCV 2.4.9.
cv2.solvePnP requires that your input object and image arrays are in contiguous memory (one of the things that the Mat::checkVector() function requires by default). Further, if you use the CV_P3P method, it relies on cv::undistortPoints, which requires the image points to be in a 2 channel type.
So solvePnP will error-fail with
World = array([[-0.5, -0.5, 3. ],
[ 0.5, -0.5, 3. ],
[ 0.5, 0.5, 3. ],
[-0.5, 0. , 3. ]])
keyPoints = array([[ 279.03286469, 139.80463604, 1. ],
[ 465.40665724, 136.70519839, 1. ],
[ 465.40665724, 325.1505936 , 1. ],
[ 279.03286469, 230.927896 , 1. ]])
objectPoints = World
imagePoints = keyPoints[:,:2] # <--- THIS SLICE IS A PROBLEM CAUSER!!!
cv2.solvePnP(objectPoints, imagePoints, np.eye(3), np.zeros(5))
but it will succeed, even in CV_P3P mode, with
imagePoints = np.ascontiguousarray(keyPoints[:,:2]).reshape((4,1,2)) # Now OpenCV is HAPPY!
retval, rvec, tvec = cv2.solvePnP(op, ip, intrinsic_mx, distortion_coefs, flags=cv2.CV_P3P)
