matchTemplate with SQDIFF_NORMED clips at 1.0

[jumostedu]

System information (version)

• OpenCV => latest master build Aug. 1 2019 (f9cefc8) -> 4.1.1-dev
• Operating System / Platform => Linux
• Compiler => gcc

Detailed description

With method SQDIFF_NORMED matchTemplate() returns 1.0 in all places where the actual result according to the formula in the documentation it should be >1.0. Notice that in contrast to the other _NORMED-methods the normalization of SQDIFF_NORMED is mathematically not guaranteed to be <=1.

Steps to reproduce

For one value:

import cv2
import numpy as np

i = np.array([[5, 1],
              [2, 8]], dtype=np.float32)
t = np.array([[1, 5],
              [7, 1]], dtype=np.float32)
m = np.ones(t.shape, dtype=t.dtype)

r = cv2.matchTemplate(i, t, cv2.TM_SQDIFF_NORMED)
print(r)

# Calculate SQDIFF_NORMED by hand
sqdiff = np.sum((i - t)**2)
norm = np.sqrt(np.sum(t**2) * np.sum(i**2))
print(sqdiff/norm)

With a larger array: you can see that there are multiple values with value exactly 1.0, so this already seems off, if you calculate the correct value at those points you see that they should be greater.

import cv2
import numpy as np

def genimg(size):
    return np.round(2*np.random.uniform(size=size).astype(np.float32))
i = genimg(42).reshape((6, 7))
t = i[:2,:2].copy()
m = np.ones(t.shape, dtype=t.dtype)
r = cv2.matchTemplate(i, t, cv2.TM_SQDIFF_NORMED)
print(r)

[asmorkalov]

opencv/opencv_extra#916