Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Failed to load latest commit information.

ALPR in Unscontrained Scenarios


This repository contains the author's implementation of ECCV 2018 paper "License Plate Detection and Recognition in Unconstrained Scenarios".

If you use results produced by our code in any publication, please cite our paper:

  author={S. M. Silva and C. R. Jung}, 
  booktitle={2018 European Conference on Computer Vision (ECCV)}, 
  title={License Plate Detection and Recognition in Unconstrained Scenarios}, 


In order to easily run the code, you must have installed the Keras framework with TensorFlow backend. The Darknet framework is self-contained in the "darknet" folder and must be compiled before running the tests. To build Darknet just type "make" in "darknet" folder:

$ cd darknet && make

The current version was tested in an Ubuntu 16.04 machine, with Keras 2.2.4, TensorFlow 1.5.0, OpenCV 2.4.9, NumPy 1.14 and Python 2.7.

Download Models

After building the Darknet framework, you must execute the "" script. This will download all the trained models:

$ bash

Running a simple test

Use the script "" to run our ALPR approach. It requires 3 arguments:

  • Input directory (-i): should contain at least 1 image in JPG or PNG format;
  • Output directory (-o): during the recognition process, many temporary files will be generated inside this directory and erased in the end. The remaining files will be related to the automatic annotated image;
  • CSV file (-c): specify an output CSV file.
$ bash && bash -i samples/test -o /tmp/output -c /tmp/output/results.csv

Training the LP detector

To train the LP detector network from scratch, or fine-tuning it for new samples, you can use the script. In folder samples/train-detector there are 3 annotated samples which are used just for demonstration purposes. To correctly reproduce our experiments, this folder must be filled with all the annotations provided in the training set, and their respective images transferred from the original datasets.

The following command can be used to train the network from scratch considering the data inside the train-detector folder:

$ mkdir models
$ python eccv models/eccv-model-scracth
$ python --model models/eccv-model-scracth --name my-trained-model --train-dir samples/train-detector --output-dir models/my-trained-model/ -op Adam -lr .001 -its 300000 -bs 64

For fine-tunning, use your model with --model option.

A word on GPU and CPU

We know that not everyone has an NVIDIA card available, and sometimes it is cumbersome to properly configure CUDA. Thus, we opted to set the Darknet makefile to use CPU as default instead of GPU to favor an easy execution for most people instead of a fast performance. Therefore, the vehicle detection and OCR will be pretty slow. If you want to accelerate them, please edit the Darknet makefile variables to use GPU.


License Plate Detection and Recognition in Unconstrained Scenarios




Unknown and 2 other licenses found

Licenses found






No releases published


No packages published