How to train two objects with different sizes in openCv [closed]
I try to train the licence plate car and licence plate motorcycle with OpenCv version 2.4.10.
I have a 235 images positives and a 300 negative images.
For positive images the licence plate car the real size is 40 cm x 13 cm and licence plate motorcycle have real size 20 cm x 17 cm. When tried to train the vectors separately the software works fine, but when I tried to train utilizing the two plates, the software doesn't work.
Here is the utilized code of training.
This code works with a licence plate car, but doesn't works with licence plate motorcycle
createsamples -info c:\harrkit\HAARKITV10\povitivos\info.txt -vec vetor\veiculos_64x24.vec -num 235 -w 64 -h 24
trainCascade -data cascade -vec c:\harrkit\HAARKITV10\vetor\veiculos_64x24.vec -bg c:\harrkit\HAARKITV10\negativos\negativos.txt -numPos 235 -numNeg 200 -numStages 12 -featureTyp HAAR -minHitRate 0.999 -maxFalseAlarmRate 0.5 -w 64 -h 24
// code detectMultiScale
if (mAbsolutePlateSize == 0) {
int height = mGray.rows();
if (Math.round(height * mRelativePlateSize) > 0) {
mAbsolutePlateSize = Math.round(height * mRelativePlateSize);
}
mNativeDetector.setMinFaceSize(mAbsolutePlateSize);
}
MatOfRect mMatPlacas = new MatOfRect();
if (mDetectorType == JAVA_DETECTOR) {
if (mJavaDetector != null)
mJavaDetector.detectMultiScale(mGray, mMatPlacas, 1.1, 2, 2, new Size(mAbsolutePlateSize, mAbsolutePlateSize), new Size());
} else if (mDetectorType == NATIVE_DETECTOR) {
if (mNativeDetector != null)
mNativeDetector.detect(mGray, mMatPlacas);
} else {
Log.e(TAG, "Detection method is not selected!");
}
That is quite normal, since the ratio of the input data is NOT identical, nor will it ever be. This means that before training, your opencv_createsamples application will force the motorcycles numberplates to a form that will never occur in test images and thus it will never be retrieved. Only way to achieve what you want is to use a double model classification.
License plate car
License plate motorcycle
My remark stays, it is NOT possible to combine them!
@Fabruccio, it is simply using two times a single model ... like you said in your own question and which is proved to work ...
Do You mean, train the objects separately?
Yes, train a single model for the car numberplates and a single model for the motorbike numberplates!