Conflicting Numpy and OpenCV2 Datatypes when calling OpenCV functions

Question

I have a big problem when using the OpenCV 2 Python API. There are no more separate OpenCV Matrix types. Every matrix is actually a numpy matrix. So far so good. The problem arises when calling OpenCV functions on these matrices that require a specific data type. OpenCV seems to have problems reconciling numpy data types with OpenCV datatypes. For instance, numpy matrices with np.uint8 don't seem to be recognized as cv_8uc1.

Here's a concrete example where it goes wrong, when trying to do a distance transform on a thresholded image:

    # threshold operation
    val, thr =  cv2.threshold(img, 64, 255, cv2.THRESH_BINARY )

    # storage matrix for the distance map
    map = np.zeros((rows,cols,1), np.uint8)

    # attempt to apply distance transform
    out  = cv2.distanceTransform(thr, cv2.DIST_LABEL_CCOMP, 3, map)

This produces the following error:

    OpenCV Error: Unsupported format or combination of formats (source 
    imagemust be 8uC1 and the distance map must be 32fC1 (or 8uC1 in 
    case of simple L1 distance transform)) in cvDistTransform
    ....
    2.4.8/modules/imgproc/src/distransform.cpp:723: error: (-210) 
    source image must be 8uC1 and the distance map must be 32fC1 
   (or 8uC1 in case of simple L1 distance transform) in function      
    cvDistTransform

thr.dtype is np.uint8 so I don't know why this error appears. Does OpenCV 2 not map from the numpy datatypes to the corresponding OpenCV data types?

I would be very glad to find a solution to this problem.

Answer 1

Could not reproduce...

import cv2
import numpy as np

thr = np.random.rand(100,100).astype(np.uint8)
map = np.zeros((100,100,1), np.uint8)
out = cv2.distanceTransform(thr, cv2.DIST_LABEL_CCOMP, 3, map)
# no errors

You could double-check used datatypes.

python 2.7.3
numpy 1.6.1
cv2 2.4.10