search

mcv-m6-video / support2016

Use the issues section of this repository to pose your question regarding the assignments.
0 stars 0 forks source link

Doubt about Stauffer & Grimson performance #2

Closed Guim3 closed 8 years ago

Guim3 commented 8 years ago

Hello,

we have been using one gaussian recursive, non recursive and Stauffer & Grimson approach to segment the background and foreground of the video sequences. We have found that S&G has given the lowest of the F1Scores (40%-50%), while the others have a much greater F1Score (60%-70%). Is that normal? We expected the S&G to be the most elaborate method and the one that would give us the best F1Scores. We are not sure if these are the results we should have obtained.

Thank you and happy new year!

Pau, Adrià and Guim

jrhupc commented 8 years ago

Hi Pau, Adrià, Guim,

What implementation are you using for S&G? the matlab one? In theory the S&G should give better results than you implementation with 1 gaussian. However, as it has several parameters to tune sometimes it is more difficult to obtain those results!

My advise is to focus on a sequence that should "favor" the S&G multiple gaussians approach. The "fall" sequence, having a non-static background (with the leaves of the tree) should be modelled better with the S&G, play a bit with the number of gaussians in the S&G parameters and you should see an improvement with respect to your implementation.

Bon any!!

Guim3 commented 8 years ago

Thanks for answering so fast!

Yes, we are using the MATLAB implementation (vision.ForegroundDetector http://es.mathworks.com/help/vision/ref/vision.foregrounddetector-class.html ). And also, we have tested with 3, 4, 5 and 6 Gaussians, but the results are quite the same for all these values (being 3 Gaussian the optimal number). Should we try to tune more parameters to improve the F1Score?

Thanks again!

jrhupc commented 8 years ago

Hello again,

Yes, I would advise to try some other parameters. Make sure than you use the same number of frames for learning and testing and that you're using the same color/gray space for comparing both algorithms.

Beat regards, El 31 dic. 2015 9:26 p. m., "Guim3" notifications@github.com escribió:

Thanks for answering so fast!

Yes, we are using the MATLAB implementation (vision.ForegroundDetector http://es.mathworks.com/help/vision/ref/vision.foregrounddetector-class.html ). And also, we have tested with 3, 4, 5 and 6 Gaussians, but the results are quite the same for all these values (being 3 Gaussian the optimal number). Should we try to tune more parameters to improve the F1Score?

Thanks again!

— Reply to this email directly or view it on GitHub https://github.com/mcv-m4-video/support/issues/2#issuecomment-168241258.