Skip to content

Latest commit

 

History

History
68 lines (51 loc) · 3.5 KB

README.md

File metadata and controls

68 lines (51 loc) · 3.5 KB
Raw

License Plate (LP) Detection

We use the CCPD dataset introduced by paper Towards End-to-End License Plate Detection and Recognition: A Large Dataset and Baseline.

Recent Update

  • 2019.10.02 model v1 for CCPD dataset is released.
  • 2020.03.27 model v1_small for CCPD dataset is released.(half parameters, much faster, much more friendly for edge devices)

Brief Introduction to Model Version

  • v1 - is designed for CCPD dataset, covering LP scale [64, 512]. It has 3 branches. Please check ./symbol_farm/symbol_structures.xlsx for details.
  • v1_small - the same as v1 but has less parameters, resulting much faster inference speed. Please check ./symbol_farm/symbol_64_512_16L_3scales_v1_small.py for details.

Inference Latency

  • Platform info: NVIDIA RTX 2080TI, CUDA 10.0, CUDNN 7.4.2, TensorRT 5.1.5.0
Model Version 320×240 640×480 1280×720 1920×1080 3840×2160 7680×4320
v1 0.62ms(1613.18FPS) 1.02ms(978.64FPS) 2.10ms(476.80FPS) 4.21ms(237.32FPS) 15.68ms(63.78FPS) 62.82ms(15.92FPS)
v1_small 0.52ms(1936.41FPS) 0.91ms(1093.00FPS) 1.70ms(586.58FPS) 3.39ms(294.72FPS) 12.12ms(82.51FPS) 47.67ms(20.98FPS)
  • Platform info: NVIDIA GTX 1060(laptop), CUDA 10.0, CUDNN 7.4.2, TensorRT 5.1.5.0
Model Version 320×240 640×480 1280×720 1920×1080 3840×2160
v1 0.86ms(1167.71FPS) 1.83ms(546.00FPS) 4.45ms(224.63FPS) 9.68ms(103.27FPS) 37.59ms(26.60FPS)
v1_small 0.67ms(1491.29FPS) 1.24ms(808.03FPS) 2.71ms(368.34FPS) 5.63ms(177.52FPS) 21.61ms(46.28FPS)

CAUTION: The latency may vary even in the same setting.

Accuracy on CCPD Dataset

We use the latest CCPD dataset, containing 351,974 images (it is larger than the version described in the paper). Since the train/test split is not provided by the paper, we randomly select 3/5 data for training and the rest is for test. We train v1 on the training set (211,180 images) and evaluate on the test set (140,794 images).

Quantitative Results on Test Set

Average Precision (AP) is used for measuring the accuracy. In detail, we use code Object-Detection-Metrics for calculating the AP metric. The following table presents the results:

The comparison is not fair due to different traning/test split. This is for reference only!

We make only one inference for each image in test. So some extremely large plates are failed to detect.

Method AP
RPnet [1] 0.945
v1 0.989
v1_small 0.982

[1] Towards End-to-End License Plate Detection and Recognition: A Large Dataset and Baseline

Some Qualitative Results on Test Set

Some challenging cases are presented.

image image image image image image image

User Instructions

Please refer to README in face_detection for details.

Data Download

Please visit CCPD for accessing the data.