python opencv cv2中的面具不起作用？

Question

While in general the new python bindings for opencv (cv2) are a beauty, "masks" don't seem to be working properly - unless I really get something wrong:

For example "cv2.add" still works properly without a mask:

import cv2
a = ones((2,2,3), dtype=uint8)
cv2.add(a,a)

correctly gives

array([[[2, 2, 2],
        [2, 2, 2]],

       [[2, 2, 2],
        [2, 2, 2]]], dtype=uint8)

But when you add a mask (and an out array "b" - which is required by for some reason is not assigned either) you get a RANDOM result, ie the result changes when you run the command multiple times

myMask = zeros(a.shape[0:2], dtype = uint8)
mask[1,1] = 255
b = zeros(a.shape)
cv2.add(a,a,b,myMask)
cv2.add(a,a,b,myMask)

gives on my machine (Win7, 32bit,Python 2.7, opencv 2.3.1)

In [34]: cv2.add(a,a,b,myMask)
Out[34]: 
array([[[ 26,   0, 143],
        [  5, 216, 245]],

       [[156,   5, 104],
        [  2,   2,   2]]], dtype=uint8)

In [35]: cv2.add(a,a,b,myMask)
Out[35]: 
array([[[35,  0,  0],
        [ 0,  3,  0]],

       [[ 0,  0,  3],
        [ 2,  2,  2]]], dtype=uint8)

... and something new on the next trial. Now either I get something seriously wrong, or there is a serious problem with the cv2 bindings.

Any suggestions?

Answer 1

Its an interesting question. I am seeing the same problem. I posted a bug and got a reply. http://code.opencv.org/issues/1748

The solution is simple. The dst array is undefined on creation and the operation changes only those destination array pixels p, for which mask(p)!=0.

So the only mechanism that works is to premake dst before addition. Ie

dst = np.zeros(...)
dst = cv2.add(a, a, dst=dst, mask=mask)

The next release will clear newly created images in operations such as cv2.add, cv2.subtract, cv2.bitwise_and/or/xor - so it will work without problem.

my code looks like:

import cv2
import numpy as np
import time

a = np.ones((2,2,3), dtype=np.uint8)

print "simple add"
t = time.time()
for i in range(10000):
    b = cv2.add(a,a)
print "%5.4f seconds" % (time.time()-t)
print b

print "\nnumpy add"
t = time.time()
for i in range(10000):
    b = a+a
print "%5.4f seconds" % (time.time()-t)
print b

# make mask same dimensions but 1 byte deep(not three)
mask = np.zeros(a.shape[:-1], dtype=np.uint8)
mask[1,1] = 255

print "\nmask", mask.shape
print mask

print "\nmasked add - uninitialised"
t = time.time()
for i in range(10000):
    b = cv2.add(a,a,mask=mask)
print "%5.4f seconds" % (time.time()-t)
print b
print "uninitialised entries are unmodified - so random.\n Inconsistent when run more than once."
print "same calc a second time..."
b = cv2.add(a,a,mask=mask)
print b

print "\nmasked add - using preinitialised dst"
t = time.time()
b = a.copy()
for i in range(10000):
    b = cv2.add(a,a,b,mask=mask)
print "%5.4f seconds" % (time.time()-t)
print b
print "Consistent when run more than once."
print "same calc a second time..."
b = a.copy()
b = cv2.add(a,a,b,mask=mask)
print b

FYI: timings (10k repeats):

cv2.add - no mask            0.0120 seconds
cv2.add - with mask          0.0160 seconds
np.add                       0.0190 seconds
cv2.add - uninitialised mask 0.0220 seconds

FYI: Submit bugs following instructions here: http://code.opencv.org/projects/OpenCV/wiki/WikiStart

Answer 2

why not use numpy (inplace) masked expression?

b = array(a, copy=True)
b[mask] += a

mask is a boolean array, analog in opencv (cv2.add) where a value is changed when mask != 0

PS: however, your code works fine in my machine (Win7 64, Python 2.6(32bit), OpenCV 2.3.0) though