View Single Post
Old 26th April 2012, 21:13   #4  |  Link
jmac698
Registered User
 
Join Date: Jan 2006
Posts: 1,867
That's really nice - good job!
It's almost what I need. It could be more flexible with just a few changes in logic.
My problem, which is particularly irksome, is I have a set of frames which are actually the same except for a) extreme noise and b) dupes in each that puts them out of sync eventually.
The problem is that due to a) there is no dupe finder or even frame matching possible, however there is frame *un*matching; when the camera is panning say the change is large enough to reliably detect.
So what I need is an ordered sort of unmatching frames. Let me be more clear with an example:
[code]
A B Match?
1 1 yes
2 1 yes (falsely because of low motion)
3 3 yes
4 5 yes (falsely, also we're out of sync now)
5 6 no (noticed beause there's camera pan now)
6 7 no
.......
31 32 yes (pan is finished, false matching again)
I know that 4-30 don't match, but I'd like to know that a shift of 1 matches this string for n frames better than another offset. So I guess a multiframe match is more robust. At some point I'd know I was in sync from n1 to n2 and out of sync from n3 to n4, and the loss was somewhere in n2-n3, which I'd have to check by hand.
I believe I explained this before and Gavino provided an answer but I didn't get back to it, to explain that I required a more robust solution.
So Id like to say, for a sync shift of up to n frames, do an ordfered search for each offset until you've eliminated an offset by virtue of it definitely mismatching for m frames. For n>=2 you could end up with two theories which aren't eliminated; then log the result with a confidence. But over enough video you hould always find areas with definite mismatch
I probably have to think about this more.

Bear in mind that when I say matching here, the tolerance is at a really high level because of all the noise.
One more thing, I'd like to see the matching function just separated out in code to return a value between two single frames, then I can put other matching functinos in there. I'd have to test some, but I like correlation because I can prove that mathematically it will match even different brightness or contrast, but not match really different pixels. That's good for my case.

Last edited by jmac698; 26th April 2012 at 21:26.
jmac698 is offline   Reply With Quote