Vietnamese Bike License Plate Recognition Challenges
- Overview
Automatic number plate recognition (ANPR) is one of the main functionalities in many automated parking and intelligent transportation systems. ANPR takes the input from a single image or a video stream and outputs the location, if exist, and the number of the plate(s). Though many commercial systems have been successfully deployed, technical challenges still exist in plate detection, character segmentation and recognition.
- Task
The challenge has two main tasks:
- Plate detection: The objective is to locate the rough location of the plate existing in the image, in terms of the bounding box that covers the plate image.
|
Input |
Output |
|
|
(x, y, width, height): 397 223 121 92
|
- Plate recognition: Given the location of the plate, the objective is to recognize the content, or the text of the plate.
|
Input |
Output |
|
(x, y, width, height): 397 223 121 92
|
52M6-9848 |
- Dataset
We prepared a set of 3,000 still images of motorbike captured at a parking slot in a hotel in Vietnam. 2,000 images will be used as a training data for developing models and the remaining 1,000 images are used for testing.
Each image file is associated with a text file containing the ground truth information. For example, the ground truth of image_0010.jpg is the image_0010.txt that contains the string
“1 397 223 121 92 52M6-9848”
That means there is one plate in the image at the location of (x, y, width, height) = (397, 223, 121, 92). The plate number is “52M6-9848”.
Link the download the training data: update soon
- Result
For the evaluation we will compare the detected plate and the recognition result with the ground truth data. The best method will be the one with the smallest error in each task.
- Organizer
VAPR - Vietnamese Association for Pattern Recognition
