We are happy to release UC3M-LP, the largest open-source dataset for European license plate detection and recognition and the first one ever dedicated to Spanish license plates.
It contains 1975 images from 2547 different vehicles, comprising a total of 12757 plate characters. It is conceived to be used mainly for character detection and OCR purposes with deep learning models. It can be directly downloaded from here
The annotations pursue a twofold objective: the plate polygonal label allows to train a model to perform a license plate rectification in a preprocessing stage and the character bounding boxes give the possibility to fine-tune an OCR model. However, to be GPDR compliant, some random characters need to be blurred in each plate and the vehicle color has been modified to preserve the anonymity.
This dataset enables the creation of many different applications, such as vehicle surveillance, traffic analysis, and security. This is the case of the proposed dual recognition system –included in the paper–. It leverages the use of OCR and visual features for robust recognition. We also used the UC3M-VRI (visual re-identification) dataset, which is also available as open-source.