Ask Your Question

What do you think about adding CMT tracker to the OpenCV?

asked 2018-11-24 15:56:41 -0500

wojtasso gravatar image

Is it good idea? I saw there is a TLD tracker, but there is no CMT.

edit retag flag offensive close merge delete


yea, nice idea !

berak gravatar imageberak ( 2018-11-25 06:12:16 -0500 )edit

So basicaly I can take the code from here, use opencv template and make a pull request to opencv_contrib?

wojtasso gravatar imagewojtasso ( 2018-11-25 06:38:34 -0500 )edit

AFAIK according to results TrackerCSRT is the best tracker implemented in contrib so far. IMHO CMT is unnecessary.

sturkmen gravatar imagesturkmen ( 2018-11-25 07:07:20 -0500 )edit

@wojtasso, no, if at all you'd have to work on the cpp version , and make it behave like the other trackers in the opencv module. (they won't accept pure python code, unless it's a demo for opencv's python wrappers)

berak gravatar imageberak ( 2018-11-25 07:15:57 -0500 )edit

so, it worked quite nice with my webcam, but bad on the benchmark

at some point, the bbox just gows HUUGE, and does no more recover. (different from my webcam experiments)

image description

on the other hand, there's like a dozen unsolved issues with the current code, things can probably be improved..

berak gravatar imageberak ( 2018-11-25 07:46:01 -0500 )edit

yes i tried it before. there are too many tracker implementations in github.for example different KCF implementations ( to be honest most of them works better than OpenCV's) . i am not against adding new algorithms but feel it is a bit unnecessary. Sorry for my laziness to explain more (with my poor english) :)

sturkmen gravatar imagesturkmen ( 2018-11-25 08:03:28 -0500 )edit

yea, left kcf out on purpose (it loses it after 100 frames there.)

and right now, i'd think: the quality should improve a lot, to be acceptable.

berak gravatar imageberak ( 2018-11-25 08:05:53 -0500 )edit

@berak sorry I pasted wrong link. Of course I'll have an intention to work on CppMT code.I think that these issues can be solved during introducing the code to the OpenCV.

@berak@sturkmen So what do you think? Should I start to work on this code? Because if this code did not pass the PR I think there is no reason to work on it.

wojtasso gravatar imagewojtasso ( 2018-11-25 08:27:34 -0500 )edit

i'm undecided. ;)

i ran the benchmark again with code from here and data from here and the other 2 videos actually give nice results for CMT. /even as it is now)

if this code did not pass the PR

don't be too afraid of that. if it's halfway decent, they'll accept it.

on the other hand, take a look at the VOT challenge , again. things are evolving very fast here !

(maybe your energy is better spent on another algorithm ?)

berak gravatar imageberak ( 2018-11-25 09:45:37 -0500 )edit

let me run benchmark later with vot2018 dataset and tell the results

sturkmen gravatar imagesturkmen ( 2018-11-25 10:10:07 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2018-12-01 06:09:00 -0500

berak gravatar image

updated 2018-12-08 01:51:27 -0500

adding newer trackers is a better idea.

MKCFup seems better according to this

also STAPLE can be easily integrated

edit flag offensive delete link more


ah, didn't even knew MKCFup, thanks ;)

(unfortunately, it has a dependancy to fftw3)

berak gravatar imageberak ( 2018-12-08 01:46:55 -0500 )edit

Question Tools

1 follower


Asked: 2018-11-24 15:55:24 -0500

Seen: 401 times

Last updated: Dec 08 '18