I wonder what the structural similarity is between two arbitrary images, then. If SS(A,-A) < SS(A,B) for some arbitrary B, then it suggests something wrong with the similarity metric.
In my previous blog, you can find how I calculate the MSSIM. I downloaded Wang's codes for measuring. You can check them too. They give some examples and details about the implementation based on their paper. http://hullsmurf.blogspot.com/2009/08/little-better-than-pangs-method-in-tone.html#comments
I don't mean that there is an error in the implementation. I mean that it may not be designed in such a way as to show similarity as a human observer would interpret similarity.
This suggests a possible research question. Can we design a "contrast similarity" metric? Contrast better captures the properties of interest for us. Perhaps there are better metrics available already -- should search the vision literature a bit.
MSSIM(A,C)=0.3243
MSSIM(A,C)=0.3243
MSSIM(A,255-A) = -0.1978
MSSIM(A,255-A) = -0.1978
MSSIM(A,A) = 1
MSSIM(A,A) = 1
Thanks, this is very interesting.
ReplyDeleteI wonder what the structural similarity is between two arbitrary images, then. If SS(A,-A) < SS(A,B) for some arbitrary B, then it suggests something wrong with the similarity metric.
In my previous blog, you can find how I calculate the MSSIM. I downloaded Wang's codes for measuring. You can check them too. They give some examples and details about the implementation based on their paper.
ReplyDeletehttp://hullsmurf.blogspot.com/2009/08/little-better-than-pangs-method-in-tone.html#comments
I don't mean that there is an error in the implementation. I mean that it may not be designed in such a way as to show similarity as a human observer would interpret similarity.
ReplyDeleteThis suggests a possible research question. Can we design a "contrast similarity" metric? Contrast better captures the properties of interest for us. Perhaps there are better metrics available already -- should search the vision literature a bit.
Ok. I will search them.
ReplyDelete