I couldnt find this elsewhere, so apologies if it's already been fixed. Applies to Linux x64, Opencv 2.3.1, svn6514

In the API, ProjectPoints2 [[ProjectPoints]]2 claims that object points can be 3xN or Nx3 (for one-channel), likewise image points can be 2xN or Nx2.

This python code, taken from https://code.ros.org/trac/opencv/browser/trunk/opencv/tests/python/test.py?rev=2296, succeeds in one case and fails in the other (commented lines).

<pre>
import cv
intrinsic_mat = cv.CreateMat(3, 3, cv.CV_32FC1);
distortion_coeffs = cv.CreateMat(1, 4, cv.CV_32FC1);

cv.SetZero(intrinsic_mat)
intrinsic_mat[0,2] = 320.0
intrinsic_mat[1,2] = 240.0
intrinsic_mat[0,0] = 320.0
intrinsic_mat[1,1] = 320.0
intrinsic_mat[2,2] = 1.0
cv.SetZero(distortion_coeffs)
distortion_coeffs[0,0] = 1e-1

fptypes = [cv.CV_32FC1, cv.CV_64FC1]
for t0 in fptypes:
for t1 in fptypes:
for t2 in fptypes:
for t3 in fptypes:
rotation_vector = cv.CreateMat(1, 3, t0)
translation_vector = cv.CreateMat(1, 3, t1)
object_points = cv.CreateMat(7, 3, t2)
image_points = cv.CreateMat(7, 2, t3)
# object_points = cv.CreateMat(3, 7, t2) # Fails
# image_points = cv.CreateMat(2, 7, t3)
cv.ProjectPoints2(object_points, rotation_vector, translation_vector, intrinsic_mat, distortion_coeffs, image_points)

</pre>