CODEWITHSUNDEEP
pythonarraysopencvnumpystride
wim
wim

Reputation: 363073

How to convert a numpy array view to opencv matrix?

I'm using opencv v2.2 to do some template matching on ndarrays, and I had great trouble with memory leaks when using their wrapped method cv.fromarray(). Rather than plug the memory leaks I avoided the fromarray() function and used cv.SetData directly, like this: 

assert foo_numpy.dtype == 'uint8'
assert foo_numpy.ndim == 3
h, w = foo_numpy.shape[:2]
foo_cv = cv.CreateMat(h, w, cv.CV_8UC3)
cv.SetData(foo_cv, foo_numpy.data, foo_numpy.strides[0])

This seems to solve the memory leaks and foo_cv seems to be deallocated properly when it goes out of scope. However, now I have the issue where if foo_numpy is just a slice/view on a bigger array, I'm not permitted foo_numpy.data (cannot get single-segment buffer for discontiguous array). At the moment I'm working around this by making foo_numpy.copy() if foo_numpy.base != None, which permits getting the buffer on the new copy. But I have the feeling this is unnecessary, the slice has the __array_struct__ and __array_interface__ so I should be able to just stride it with the appropriate stepsizes somehow? I'm not sure how to do it in a nice way, because the base of this one can also be a view on another larger array ad infinitum.

Upvotes: 4

Views: 4205

Answers (1)

rroowwllaanndd
rroowwllaanndd

Reputation: 3958

I think the problem with what you were trying to do is that the array data you're interested in (ie. foo_np_view) is actually only stored in one place i.e. foo_np.data, and the OpenCV SetData method doesn't provide any way to specify stride settings that would allow you to skip the bytes that are not part of foo_np_view.

You can, however, get around this problem using Numpy’s tostring() method, which turns an array (or views therein) into a byte string:

>>> import numpy as np
>>> import cv
>>> foo_np = np.array( 255 * np.random.rand( 200 , 300 , 3 ), dtype = 'uint8' )
>>> foo_np_view = foo_np [ 50:150:2 , 10:290:5 , : ]
>>> h,w,d = foo_np_view.shape
>>> foo_cv = cv.CreateMat( h , w , cv.CV_8UC3 )

Recreating the original problem:

>>> cv.SetData( foo_cv , foo_np_view.data, foo_np_view.strides[0] )
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: cannot get single-segment buffer for discontiguous array

Using the tostring() method (see below for explanation of the stride setting):

>>> cv.SetData( foo_cv , foo_np_view.tostring() , w * d * foo_np_view.dtype.itemsize )
>>> np.array_equal( np.asarray( foo_cv ) , foo_np_view )
True

The value w * d * foo_np_view.dtype.itemsize gives us a stride value identical to that of foo_np_view.copy(), which is necessary as the string representations of the view and its copy are identical:

>>> foo_np_view.copy().tostring() == foo_np_view.tostring()
True
>>> foo_np_view.copy().strides[0] == w * d * foo_np_view.dtype.itemsize
True

Upvotes: 2

Related Questions