What it does Trains Yolov5 to detect Chinese license plates across eight specific types: blue, yellow, green (new energy and agricultural), black, white (police, military, armed police), and double-decker variants. Ships with both PyTorch inference (detect_plate.py) and an NCNN C++ project for edge deployment.

The interesting bit This is the fifth architectural iteration in a long lineage of plate detectors from the same group. They started with classical computer vision (LBP/Haar), graduated through MobileNet-SSD and RetinaFace, and finally landed on Yolov5—explicitly because earlier versions couldn’t keep up on speed or accuracy. The training data mixes the public CCPD dataset with their own private collection, which is how they cover rarer plate styles like double-decker military plates.

Key highlights

• Supports 8 distinct Chinese plate types, including niche categories like civil aviation and agricultural green plates
• Dual runtimes: Python/PyTorch 1.7 for prototyping, C++/NCNN for production edge inference
• Trained on CCPD + proprietary data; claims broader plate-type coverage than CCPD alone
• Directly references yolov5-face as architectural inspiration
• 1,287 stars suggests modest but real adoption in the Chinese CV community

Caveats

• README is entirely in Chinese; no English documentation
• PyTorch dependency pinned to 1.7.0 and Python 3.8, which are increasingly dated
• No quantitative benchmarks, speed numbers, or accuracy metrics provided—just a single result image
• “Faster and more accurate” claims are stated but not measured against prior versions or competitors

Verdict Worth a look if you’re building a Chinese traffic or parking system and need off-the-shelf plate localization with NCNN deployability. Skip it if you need OCR (this detects bounding boxes only), non-Chinese plates, or reproducible benchmarks to compare against your own model.