I try to replace my 'home-brew' gaussian mixture model with the ipp version, but I run into two problems:
- Setting the parameters of the model doesn't seem to have any influence on the results
- The reference image buffer doesn't get filled at all.
- I allocate a IppFGGaussianModel and set it's parameters.
- I call ippiForegroundGaussianInitAlloc
- and for each incoming image I call ippiForegroundGaussian
Am I doing something wrong ? Or is this a bug ?
I can see absolutely no change in the reference image no matter what parameters are set. The output image does produce output, but seems to behave very strange. Noisy, as Rob said would be a good (if gentle) discription.
Some guidance on proper usage of these functions would be appreciated.
From the latest release notes:
'Corrected results from calling Image Segmentation function ippiForegroundGaussian'
So I ran my test project using the new version:
I must admit the reference image does get filled now. The segmentation result however remains 'suboptimal', and changing the model parameters still doesn't affect the result at all. I suggest some more bug fixing.
I posed a similar question over a year ago 4-13-07 and never recieved an answer.
I've had to use my own code as well.
I'll look forward to improvement in the future.
PS. I love IPP but some of the functions in Chapt 14 "Computer Vision" are similar to ippiForegroundHistogram in their limitations.
I submitted this issue on premier support. And after some intial checks ( did I really use the right version etc. ) they looked into the matter. After a month I received an answer: 'the issue was added to an engineering database as a feature request'. I fear this can also be quoted as 'the issue has been send into oblivion' but I might be wrong. I'll see. I have no idea if they reproduced the problem or if they even tried.
Well I don't agree. A decent backgroundsubtraction method, like a gaussian mixture, running on IPP speed would certainly be helpfull. I would leave more time for high-level algorithms. I would prefer the Zivkovic variant over the promised one. I do agree that the current version which doesn't work at all should be removed, this way intel wastes the time of people trying to use the function only to find that it doesn't work...
p.s. Z.Zivkovic: Improved Adaptive Gaussian Mixture Model for Background Subtraction. ICPR, 2004
thanks all for your feedback on this thread. We actually do consider computer vision IPP functionality (or might be better to call it image analysis functionality?) as an important part of product. Unfortunately we not always can provide new product features at a speed anyone want, especially in case of complex functions like one discussed here.
I would suggest to create a separate thread for discussing a new features requests in computer vision library (this should make it easier to solidify all inputs). Whatalso interesting for us is to know what functionality in computer vision is really in use and what might be designed not that well and might be deprecated in future?