license plate detection using YOLO

Question

I am trying to train a model for detecting license plates of pakistani cars. I found a faster technique called YOLO. Here is a link YOLOv2

Blog which I am following to train YOLOv2 is this blog

According to this blog I need to have images of cars and I need to annotate these images (need to mark position of license plate) for preparation of test data and training data. Problem is that I already have training data of the form

and this tutorial requires me to do annotation from cars images like this.

If someone has worked with YOLO kindly tell me how can I avoid annotation and use my own training data to train a YOLO model.

If all of your training data is of the above format then probably YOLO is not the right choice for you. Yolo is object detection algorithm your training data is of the form of image classification problem. What is your test data? Images of vehicles or only images of plates? — Scorpio
I am doing detection not classification. Input to this model would be an image of car having a license plate on it. Test data contains images of vehicles. — Abdul Wahab
If the input is a car then why is the training data not car (or any other vehicle) ? Why is the training data just complete pics of number plate? — Scorpio
I think to achieve it you would have to provide the image license plates in all similar views that the Camera will caught them on the images. Giving it Big Plan images will not help at all. — KeitelDOG

Arjun Kava Arjun Kava · Accepted Answer · 2018-06-29T13:49:00

Yolo training requires following format of annotation

[class] [X] [Y] [W] [H]

i.e

0 0.443359 0.434722 0.391406 0.869444

If you have same annotation dataset, Congratulations you can start training. If you don't have same you can convert using tools available on github.

Update: Annotation is calculated from center For Example. If you are working with x1,y1,x2,y2 format you need to convert it.

def convert_to_yolo_format(path, x1, y1, x2, y2):
    """
    Definition: Converts (x1, y1, x1, y2) format to
        (x, y, width, height) normalized YOLO format.
    """
    size = get_img_shape(path) # size array [width,height]
    dw = 1./size[0]
    dh = 1./size[1]
    x = (x1 + x2)/2.0 # centroid x
    y = (y1 + y2)/2.0 # centroid y
    w = x2 - x1 # width
    h = y1 - y2 # height
    x = abs(x*dw) # divide by width
    w = abs(w*dw) # divide by width
    y = abs(y*dh) # divide by height
    h = abs(h*dh) # divide by height
    return (x,y,w,h)

license plate detection using YOLO

2 Answers