Standard Haar classifier for Russian reg plates perfors poorly

edit

You can find more information with the corresponding pull requests:

• for haarcascade_licence_plate_rus_16stages.xml: Added Haar cascade for russian cars licence plate detection #2777
• haarcascade_russian_plate_number.xml: Detector for russian vehicle registration plate #2746 and the blog post from the author: Разработка → Не соблаговолите ли больше не нарушать, сэр (мэм)?

You can find the training size by looking directly into the file:

• for haarcascade_russian_plate_number.xml it should be 60x20 (width x height)
• for haarcascade_licence_plate_rus_16stages.xml it should be 64x16

Object detection field is currently led and traditional object detection methods are outperformed by deep neural network fields. If you are concerned about precision and performance, you should look at these topics in my opinion.

Otherwise, another option is to train using your own training data and parameters.