search

zjykzj / LPDet

License Plate Detection and Recognition System
Apache License 2.0
1 stars 1 forks source link

european plates #3

Open Baxulio opened 5 months ago

Baxulio commented 5 months ago

Hello dear maintainer,

Is it possible to train and recognize european plates (only latin characters) with LPDet? Or it is only for Chinese characters?

Thank you!

zjykzj commented 5 months ago

Hello dear maintainer,

Is it possible to train and recognize european plates (only latin characters) with LPDet? Or it is only for Chinese characters?

Thank you!

At present, there is no additional dataset training plan. If you want to apply LPDet to other license plate datasets, you can try implementing it. Welcome to PR

Baxulio commented 5 months ago

Thank you for your reply and the great project with its simple, yet complete documentation.

Regarding training I have some questions, I would appreciate if you could answer them:

  1. So if I prepare my own dataset of EU plates similar to CCPD structure, LPDet should work (detection+recognition)?

  2. What do you think, which method is better (in terms of speed and accuracy): wR2 + RPNet or YOLO8+CRNN-CTC for detecting and recognition license plates? Your opinion!

Thank you!

zjykzj commented 5 months ago

Thank you for your reply and the great project with its simple, yet complete documentation.

Regarding training I have some questions, I would appreciate if you could answer them:

  1. So if I prepare my own dataset of EU plates similar to CCPD structure, LPDet should work (detection+recognition)?
  2. What do you think, which method is better (in terms of speed and accuracy): wR2 + RPNet or YOLO8+CRNN-CTC for detecting and recognition license plates? Your opinion!

Thank you!

Yes, LPDet can be applied to different formats of license plate datasets. Note that some code needs to be modified, such as license plate length, license plate labels, and so on.

In addition, LPDet refers to the ECCV 2018 paper, so compared to YOLOv8, the latter must have better recognition speed.

Baxulio commented 5 months ago

Thank you for your answer!

For testing I tried pretrained model from: https://github.com/zjykzj/LPDet/releases/download/v0.3.0/RPNet-e100.pth

But unfortunately I could recognize only one LP out of 8.

Here are these images that I predicted: 14_rpnet 13_rpnet 12_rpnet 11_rpnet 10_rpnet 9_rpnet 8_rpnet 7_rpnet

What do you think, what can be the cause?

zjykzj commented 5 months ago

Thank you for your answer!

For testing I tried pretrained model from: https://github.com/zjykzj/LPDet/releases/download/v0.3.0/RPNet-e100.pth

But unfortunately I could recognize only one LP out of 8.

Here are these images that I predicted: 14_rpnet 13_rpnet 12_rpnet 11_rpnet 10_rpnet 9_rpnet 8_rpnet 7_rpnet

What do you think, what can be the cause?

It seems that it is because wR2 cannot accurately locate the license plate position. May I ask which dataset these data come from? You can add training on this data or try better license plate localization algorithms, such as yolov5/yolov8.

Baxulio commented 5 months ago

Greetings!

Sure, I got them from internet, randomly. So you mean, I can use yolov8 to detect LP and then recognize via RPNet. Then it will be two-forward operation, right? Or is it also possible to make one-forward for yolov8+RPNet, like with wR2?

Thank you!

zjykzj commented 5 months ago

Greetings!

Sure, I got them from internet, randomly. So you mean, I can use yolov8 to detect LP and then recognize via RPNet. Then it will be two-forward operation, right? Or is it also possible to make one-forward for yolov8+RPNet, like with wR2?

Thank you!

YES, Detecting and recognizing license plates can be split into two behaviors. If you want to integrate it together, you need to do some work to delve into code modification and integration. Go ahead and do it!!!