Two identical image diff significant about image-match

[chekun]

@rhsimplex Nice, love you guys.

[chekun]

@rhsimplex vrde's fork fixed my problem, issue solved 😃

[rhsimplex]

Nope, it should return 0.0. I'll try to replicate this Monday and see what's going on.

Thanks for reporting the issue.

[rhsimplex]

Ok, I only get a distance of 0.09, which should be under the threshold for matching. However these images should be identical.

In [1]: from image_match.goldberg import ImageSignature

In [2]: gis = ImageSignature()

In [3]: path1 = 'https://cloud.githubusercontent.com/assets/1967804/18188884/53e77c7a-70e8-11e6-8e27-98c196f1e242.jpg'

In [4]: path2 = 'https://cloud.githubusercontent.com/assets/1967804/18188887/5bbba7aa-70e8-11e6-93b9-3e881bc03e66.jpg'

In [5]: sig1 = gis.generate_signature(path1)

In [6]: sig2 = gis.generate_signature(path2)

In [7]: gis.normalized_distance(sig1, sig2)
Out[7]: 0.094653959692538772

What's going on? I think one of these images is greyscale, and one is color.

In [8]: from skimage.io import imread

In [9]: imread(path1).shape
Out[9]: (1334, 750, 3)

In [10]: imread(path2).shape
Out[10]: (1334, 750)

So skimage (which image-match uses for loading the images) converts to greyscale from color with range (0,1) but if the image is already greyscale over 8 bits, it leaves it!

In [12]: imread(path1, as_grey=True)
Out[12]: 
array([[ 1.,  1.,  1., ...,  1.,  1.,  1.],
       [ 1.,  1.,  1., ...,  1.,  1.,  1.],
       [ 1.,  1.,  1., ...,  1.,  1.,  1.],
       ..., 
       [ 1.,  1.,  1., ...,  1.,  1.,  1.],
       [ 1.,  1.,  1., ...,  1.,  1.,  1.],
       [ 1.,  1.,  1., ...,  1.,  1.,  1.]])

In [13]: imread(path2, as_grey=True)
Out[13]: 
array([[255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255],
       ..., 
       [255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255]], dtype=uint8)

Hence the slightly different signatures. The correct fix would be get rid of the skimage dependency and use PIL directly. A quick fix might be to detect if an image is uint8 and divide by 255.

[rhsimplex]

Interestingly, when I try dividing by 255, I get the same distance value as you: 0.70823708184882128. Maybe something has changed with skimage. Can you tell me the output of:

import skimage
print(skimage.__version__)

?

I'm using version 0.12.3

[chekun]

@rhsimplex , Thank you for your reply.

I use pavlov/match docker image pavlov/match@9b7df7ecc867.

and the skimage version is exactly 0.12.3

Would use PIL directly fix the problem? Any plan to replace it ?

[rhsimplex]

If you're using pavlov's match, then the real problem is that it's using an out-of-date image-match build. Their docker file has the line:

pip install git+https://github.com/ascribe/[email protected]

And we're at version 1+ now. My colleague @vrde has actually ported match to use the latest version of image-match (which now uses python3), you can find that fork here: https://github.com/vrde/match. I'll ask him to make a PR against the original repository.

Sorry about the confusion, but see if using this version helps. If so, I probably won't make any changes because 0.09 is well below the threshold of what should be considered a match.

[chekun]

@rhsimplex Good news, I will use @vrde's fork and see if it woks out. Thank you very much. 👍

[rhsimplex]

You're welcome. Please let me know if it fixes your problem.

[rhsimplex]

there is now a PR dsys/match#8 on Pavlov's match. When the merge it you can pull from there directly =)