3

I have 3 numpy arrays of shape > (500, 500). I am trying to iterate over them simultaneously. I have tried two different methods but both of them are slow.

Here Ix_Ix_blur, Ix_Iy_blur and Iy_Iy_blur are of the same size. I'm trying to find features and draw it on OpenCV image.

Method 1:

for i in xrange (Ix_Ix_blur.shape[1]):
    for j in xrange(Ix_Ix_blur.shape[0]):
        A = np.array([ [Ix_Ix_blur[j][i], Ix_Iy_blur[j][i]], 
            [Ix_Iy_blur[j][i], Iy_Iy_blur[j][i]] ])
        detA = (A[0][0]*A[1][1])-(A[0][1]*A[1][0])
        traceA = A[0][0]+A[1][1]

        harmonic_mean = detA/traceA
        if(harmonic_mean > thresh):
            cv2.circle(img, (i,j), 1, (0, 0, 255), -1, 8)

This takes around 7 seconds for image of size of 512*512

Method 2:

Ix_Iy_blur_iter = np.nditer(Ix_Iy_blur)
Iy_Iy_blur_iter = np.nditer(Iy_Iy_blur)
Ix_Ix_blur_iter = np.nditer(Ix_Ix_blur)

while(not Ix_Iy_blur_iter.finished):
    try:
        A = np.array([[Ix_Ix_blur_iter.next(), Ix_Iy_blur_iter.next()],[Ix_Iy_blur_iter.value, Iy_Iy_blur_iter.next()]])
    except StopIteration:
        break
    detA = (A[0][0]*A[1][1])-(A[0][1]*A[1][0])
    traceA = A[0][0]+A[1][1]

    harmonic_mean = detA/traceA
    if(harmonic_mean > thresh):
        i = Ix_Ix_blur_iter.iterindex/Ix.shape[0]
        j = Ix_Ix_blur_iter.iterindex - Ix.shape[0]*i
        cv2.circle(img, (j,i), 1, (0, 0, 255), -1, 8)

This method also seems to take 7 seconds to iterate over the same size of image.

Is there any other way using which I can reduce the time required for iterations?

Configuration:

  • Ubuntu 12.04
  • 3rd Gen core i5 processor
  • 4 GB RAM
  • 2 GB ATI RADEON GPU (which I have turned off)
Froyo
  • 17,947
  • 8
  • 45
  • 73

1 Answers1

4

First you can use Ix_Ix_blur[j, i] instead of Ix_Ix_blur[j][i]. Ix_Ix_blur[j][i] will create a temporary array which is very slow.

To speedup element access with ndarray, you can use item() method, which return python native numeric values, and you don't need to create a temporary array A. Calculation with native numeric values is faster than numpy scalars.

for i in xrange (Ix_Ix_blur.shape[1]):
    for j in xrange(Ix_Ix_blur.shape[0]):
        a, b, c = Ix_Ix_blur.item(j, i), Ix_Iy_blur.item(j, i), Iy_Iy_blur.item(j, i)
        detA = a*c - b*b
        traceA = a + c
        harmonic_mean = detA/traceA
        if harmonic_mean > thresh:
            cv2.circle(img, (i,j), 1, (0, 0, 255), -1, 8)

For your particular problem, it's not necessary to do the calculation in a loop, you can:

detA = Ix_Ix_blur * Iy_Iy_blur - Ix_Iy_blur**2
traceA = Ix_Ix_blur + Iy_Iy_blur
harmonic_mean = detA / traceA
for j, i in np.argwhere(harmonic_mean > thresh):
    cv2.circle(img, (i,j), 1, (0, 0, 255), -1, 8)
HYRY
  • 94,853
  • 25
  • 187
  • 187