Add LPD-YuNet evaluation script on CCPD - #317

Adds an evaluation script for LPD-YuNet on the public CCPD2019 test set.

Related to #119 (License Plate Detection / LPD-YuNet / Need Contribution).

Claimed accuracy

LPD-YuNet’s model card and the original contributor (watrix.ai) do not publish a claimed number. Maintainer guidance on #119 was to pick a public Chinese-plate dataset and report OpenCV DNN accuracy.

This PR uses the official CCPD detection protocol (Xu et al., ECCV 2018, https://github.com/detectRecog/CCPD):

• each image has one plate
• the detector outputs one box (highest score)
• a box is correct iff IoU > 0.7
• report precision on DB, Blur, FN, Rotate, Tilt, Challenge and on their union (Overall; the paper labels this column AP)
• ccpd_base is train/val and is not scored

Paper baselines (CCPD detection table, IoU > 0.7):

Measured accuracy (OpenCV DNN)

Numbers will be added after a full run on CCPD2019 official test subsets.