Hi All,
I'm trying to create an xml for detecting the cars using opencv .
Now my question is that for detecting the car should i need to create 4 different xml i.e,(front,back ,left and right ) view xml's by using 4 different sets of the positive images . or else all sets of data together. and which method haartraining or traincascade is the best way to create cascade xml for object detection ??
Thank You.
When it comes to car detection normally people divide the task in front, back and lateral views, no need to train right and left separately. You should use traincascade instead of haartraining. It is the newest version of casade training application, it is faster and has more training options.
2 negatives for each positive should be a good start
thanks once again.. can u suggest good tutorial for generating classifier using traincascade . I have followed this tut http://mehebuba.blogspot.in/2013/07/cascaded-classifier-with-opencv.html done up to creating the .vec file and when i am trying to start the training for +ive and -ive . It is getting aborted. can you please help me out.
This is a pretty popular tutorial http://note.sonots.com/SciSoftware/haartraining.html . It uses the old haartraining app instead of traincascade, but the principles are the same.
Hi , i am done with creating .vec file and now when i am making for training using cascadetraining using the following command: C:\haar>traincascade.exe -data result -vec data\positive.vec -bg Negative\negative.txt -numPos 100 -numNeg 300 -numStage 10 -featureType LBP -precalcvalBufSize 1024 -precalcIdBufSize 1024
i am getting this error . OpenCV Error: Parsing error (result/params.xml(8): Duplicated key) in unknown function, file ......\src\opencv\modules\core\src\persistence.cpp, line 717 can you help me this thanks
I'm not sure what might be the problem, but you are not specifying the width and height of the positive samples. You should add two aditional parameters -w and -h
You'll have to be more specific in what you are trying and what problem is occurring.
For multi-view car detection, I suggest taking a look at other frameworks like deformable parts models (of which a latentSVM implementation is available in openCV). It has 3 views, each with their responsible HOG model existing of a root filter and several part filters.
A simple VJ cascade classifier won't do the job unless you combine multiple single models, which will be slower than a DPM model.
Well i think you need to focus on detection using feature extraction algorithms like SURF/SIFT with SVM approach , this may helpfull too for you as so also need to dataset for SVM , you can generate xml file out of it , you can use SVM approach for training all the views of vehicle or you can generate different xml files of each view. You can get the tutorials from here and here
Actually feature detection like sift and surf will probably only work if you want to detect that exact car type, The class of cars however has so much variation that there is a reason why academics don't use this approach for car and pedestrian detection.
@StevenPuttemans Thanks for my correction , but what if we have huge data of many types of cars and we train it using SVM with SURF ?
My guess you will overtrain to your data or not be able to create a general class model that seperates positives and negatives rather well. I think for your option, DPM models or combined cascade models are the way to go.
I do agree that it can work, but won't be the optimal solution. Did you try running this on live data and reach real time performance? I guess this is close to impossible with this technique :)
Asked: 2014-01-30 01:38:50 -0600
Seen: 1,076 times
Last updated: Feb 22 '14
