Reading license plates in the wild, very slowly by default

What it does

This is the author’s implementation of an ECCV 2018 paper that detects and recognizes license plates in unconstrained scenarios—think skewed angles, bad lighting, and other real-world mess. It chains together a Darknet-based vehicle detector and OCR with a Keras/TensorFlow license-plate detector to find plates and read their characters. The pipeline outputs annotated images and a CSV file, and you can retrain the plate detector on your own annotated data.

The interesting bit

Most computer vision repos from 2018 assume an NVIDIA card and saintly patience for CUDA; this one deliberately chooses CPU execution in Darknet to favor easy setup over raw speed. That hardware agnosticism comes at a cost: vehicle detection and OCR are described as “pretty slow” unless you opt into GPU support via the Darknet build configuration.

Key highlights

• Reference implementation of the ECCV 2018 paper “License Plate Detection and Recognition in Unconstrained Scenarios.”
• Bundles its own Darknet framework for vehicle detection and OCR alongside a Keras/TensorFlow plate-detector network.
• Supports training the LP detector from scratch or fine-tuning it on new annotated samples.
• GPU support is opt-in via the Darknet build configuration; CPU is the default to avoid CUDA setup.

Caveats

• The tested dependency stack is frozen in time: Ubuntu 16.04, Python 2.7, TensorFlow 1.5.0, Keras 2.2.4, and OpenCV 2.4.9.
• Vehicle detection and OCR are “pretty slow” on the default CPU setting.
• The bundled training samples contain only three annotated images; reproducing the full paper requires sourcing the original datasets.

Verdict

Worth a look if you need to reproduce the ECCV 2018 results or want an ALPR baseline that does not immediately demand a GPU and modern CUDA stack. Skip it if you are looking for a modern, production-ready system with up-to-date dependencies and out-of-the-box speed.