Following the guidelines on Github, I have managed to retrain the model with my own dataset and successfully run the retrained model on my PC using the LPR pipelines. However, the output video displayed was not the same as the example given which does not show up the license plate number on the top right.
I think this has something to do with the wrong dataset used to retrain the model. May I know how can I prepare the correct dataset to train the vehicles detection model, license plate detection model, and also LPRNet model. Thanks.
Besides that, I also tried to use the pre-trained model to run with different video source, and the LPR is not working, only the vehicles are detected. Can I know what is the requirements for the LPR application to work?
Let me clarify a couple points about the LPR models:
Regarding our LPR model - it’s specifically trained on Israeli license plates, which are yellow and contain only numbers. To get accurate results, you’ll need to test it with Israeli plates.
For your model and the detection issues you’re experiencing - could you share more details with us? Specifically, it would be helpful to see:
The postprocessing code that draws the bounding boxes
Information about your model and its training data
Sample images of the license plates it’s meant to detect
This information will help us better understand and address the issues you’re seeing.
I have tested the LPR application with another Israel road traffic from Youtube, and the application managed to detect and printed out some of the car plates.
As you can see from the image above, although there is some car plates got recognised but most of them are being ignored. May I know why it is acting like this? Is it because of the video resolution is not clear enough, or the camera angle is not correct?
Besides that, I would also like to know if I would want to run the application for other country’s vehicles, do I just need to retrain the LPRNet model, or the tiny_yolov4_license_plates model would be required too?
Other than testing with the video, I also tried to run the application with a web cam (EMeet FHD 1080p C960), however, my input camera framerate and resolution seems to be very low, so the live video is abit laggy and blur. Could you suggest some USB Camera which is good enough for the real application?
As for the replies to your questions:
I am just using the same postprocessing code provided by Hailo
For now, I am just using the pretrained model provided for the quick demo. But previously I also tried to train my own model with some random dataset found from kaggle, I think the number plates are Taiwan number plates so I will try to rerun my retrained model with Taiwan traffic video.
