Performance comparison of OpenCV-Python interfaces, cv and cv2

computer-vision, opencv, performance, python

Solution

The image returned by cv2.imread() is an array object of NumPy. So you can use NumPy's functions to speedup calculation.

The following program shows how to speedup your origin for loop version by using item(), itemset() method of ndarray object.

import time
import numpy as np
import cv2

gray = cv2.imread('lena_full.jpg',0)
height, width = gray.shape
h = np.empty((height,width,3), np.uint8)

t = time.time()
for i in xrange(height):
    for j in xrange(width):
        k = gray.item(i, j)
        if k == 127:
            h.itemset(i, j, 0, 255)
            h.itemset(i, j, 1, 255)
            h.itemset(i, j, 2, 255)
        elif k > 127:
            h.itemset(i, j, 0, 0)
            h.itemset(i, j, 1, 0)
            h.itemset(i, j, 2, 255-k)
        else:
            h.itemset(i, j, 0, k)
            h.itemset(i, j, 1, 0)
            h.itemset(i, j, 2, 0)
print time.time()-t

And the following program show how to create the palette first, and use NumPy's array index to get the result:

t = time.time()
palette = []
for i in xrange(256):
    if i == 127:
        palette.append((255, 255, 255))
    elif i > 127:
        palette.append((0,0,255-i))
    else:
        palette.append((i, 0, 0))
palette = np.array(palette, np.uint8)

h2 = palette[gray]

print time.time() - t

print np.all(h==h2)

The output is:

0.453000068665
0.0309998989105
True

The cv version output is :

0.468999862671

Note: the length of axis 0 is the height of the image, the length of axis 1 is the width of the image

Problem

A few days back, I started using new OpenCV-Python interface, `cv2`. My question is regarding the comparison of `cv` and `cv2` interface. Regarding the ease of use, new `cv2` interface has improved far greater, and it is really easy and fun to work with `cv2`. But what about speed? I made two small code snipplets, one in `cv` and another in `cv2`, to check the performances. Both does the same function, access pixels of an image, test it, make some modifications, etc. Below is the code: `cv2 interface`: ``` import time import numpy as np import cv2 gray = cv2.imread('sir.jpg',0) width = gray.shape[0] height = gray.shape[1] h = np.empty([width,height,3]) t = time.time() for i in xrange(width): for j in xrange(height): if gray[i,j]==127: h[i,j]=[255,255,255] elif gray[i,j]>127: h[i,j]=[0,0,255-gray[i,j]] else: h[i,j]=[gray[i,j],0,0] t2 = time.time()-t print "time taken = ",t2 ``` ===================================================== And result is: time taken = 14.4029130936 ====================================================== cv interface: ``` import cv,time gray = cv.LoadImage('sir.jpg',0) h = cv.CreateImage(cv.GetSize(gray),8,3) t=time.time() for i in xrange(gray.width): for j in xrange(gray.height): k = cv.Get2D(gray,j,i)[0] if k==127: cv.Set2D(h,j,i,(255,255,255)) elif k>127: cv.Set2D(h,j,i,(0,0,255-k)) else: cv.Set2D(h,j,i,(k,0,0)) t2 = time.time()-t print "time taken = ",t2 cv.ShowImage('img',h) cv.WaitKey(0) ``` ====================================================== The result is: time taken = 1.16368889809 ======================================================= See, here old `cv` is about `12 times faster` than `cv2`. And resulting images are same. (input image is of size 720x540) Why does this happen? Is cv2 slower compared to cv? Or am I making any mistake here? Is there a faster method in cv2 for the above code?

Original source