Ask Your Question

What would be the best detectors for these objects

asked 2013-08-21 14:02:47 -0500

TimeManx gravatar image

I started out with detecting people using the samples. HOG has good accuracy but a very slow for use on an android phone. Haar cascade was faster than HOG but accuracy was quite low. Also, I remember reading that Haar isn't suitable for human detection. For vehicle detection, Haar is working fine.

So could you guys suggest what would be best for detection of these objects

  • People
  • Vehicles
  • Trees
  • Buildings
edit retag flag offensive close merge delete


I'ts hard to tell which detector would be best for this object, but I suggest you try also the latentSVM detector (detection by parts).

GilLevi gravatar imageGilLevi ( 2013-08-22 03:28:16 -0500 )edit

@GilLevi I read in a post that for latentSVM detector, openCV provides classifiers for people, cars, etc. But I could only find a classifier for cats in the OpenCV folders.

TimeManx gravatar imageTimeManx ( 2013-08-22 03:54:53 -0500 )edit

Don't bother. Found it here.

TimeManx gravatar imageTimeManx ( 2013-08-22 03:59:38 -0500 )edit

2 answers

Sort by ยป oldest newest most voted

answered 2013-08-22 04:24:19 -0500

I'ts hard to tell without testing which detector would be best for these objects, but I suggest you try also the latentSVM detector (detection by parts).

Latentsvm detector had trained models for "person" and "vehicle" (buses, cars) and there is the option to train new detectors (though it's not trivial and includes using the Matlab package and translating the models to xml).

edit flag offensive delete link more


I've implemented the latentSVM detector in my android application acoording to how it was done in the latentsvm_multidetect.cpp sample but I just get a black screen with the log saying "Preview frame received. Frame size:152064" Any ideas?

TimeManx gravatar imageTimeManx ( 2013-08-22 15:57:42 -0500 )edit

Sorry, I'm not familiar with development for android .

GilLevi gravatar imageGilLevi ( 2013-08-22 16:08:33 -0500 )edit

answered 2013-08-23 13:26:04 -0500

FLY gravatar image

I think for detection of vehicles and buildings i suggest you use SURF/SIFT/BOW with FLANN BASED MATCHER and train them with SVM , there may be less accuracy but they are fast and their accuracy level is not much bad , Also look at the backgroundsubtraction algorithms , it also help you in better detection

edit flag offensive delete link more


You mean background subtraction followed by blob detection?

TimeManx gravatar imageTimeManx ( 2013-08-23 16:13:00 -0500 )edit

Also I'm a little confused about SIFT/SURF/ORB. In another post, someone suggested that feature detection techniques are more useful when we need to detect some specific images against a large database and that they aren't really useful in these kinds of scenarios.

TimeManx gravatar imageTimeManx ( 2013-08-23 16:26:55 -0500 )edit

@TimeManx yes , i think you already made it specific by mentioning people and vehicles ,and yes your first step is not delete the background and only select the object than detect the keypoints and discripstors and than you train them with BOW and SVM , its may not give you the accurate result but it will give you somehow better result , you can also you latent svm approach

FLY gravatar imageFLY ( 2013-08-24 02:11:12 -0500 )edit

I think that BOW + SVM + sliding window/pyramid can also be a used for detection, but I'm not sure if FLANN is relevant. Can you please explain?

GilLevi gravatar imageGilLevi ( 2013-08-24 04:16:48 -0500 )edit

Question Tools

1 follower


Asked: 2013-08-21 14:02:47 -0500

Seen: 1,678 times

Last updated: Aug 23 '13