I'm using the OpenCV latent SVM detection with the sample code openCV and self-converted bicycle.xml to test a single image bike.img, which has a size of about 500x300. It takes ~11 seconds to complete this detection on my laptop.
Is it too slow? Has anybody tried this? Or am I making something wrong.
I can send the code, img and model files if you would like to help. Thank you very much.
LatentSVM detector as implemented in OpenCV is still a slower version then the actual version Pedro Felzenszwalb (the inventor of the technique) posts available on his website.
He also places a C++ based code on his project website: http://people.cs.uchicago.edu/~rbg/latent/
LatentSVM isn't fast, thats a fact, but it is damn acurate if you train it good. Making it faster has been a research topic in computer vision labs around the world for some time. People have reached higher speeds by implementing it on GPU and such. You should check out the internet for some papers.
About the actual implementation in OpenCV, to make it faster, some things can be done. --> Give a size range, in stead of detecting at all possible sizes, if you know the size +- then you can use this knowledge to speed it up. --> Try to define scale regions in the images by using the knowledge of your training set.
Cheers
Asked: 2013-03-07 21:32:56 -0600
Seen: 1,172 times
Last updated: Mar 08 '13
