search

sergiomsilva / alpr-unconstrained

License Plate Detection and Recognition in Unconstrained Scenarios
Other
1.72k stars 605 forks source link

how to train OCR model? #19

Open linzhi123 opened 5 years ago

linzhi123 commented 5 years ago

Hello, do you have open source training OCR model code?

sergiomsilva commented 5 years ago

Hi @linzhi123 , sorry but we do not plan to publish this part anytime soon.

danishansari commented 5 years ago

@sergiomsilva Aren't you just using Darknet for OCR? If so, OCR can be trained separately using yolo, am i correct?

linzhi123 commented 5 years ago

Yeah, I just want to train OCR alone. There's no training code. Yolo's training code is official, and his training code is modified. @danishansari

linzhi123 commented 5 years ago

Do you have training code? Or do you know how to train? @danishansari

danishansari commented 5 years ago

Yes i am training it using YOLO-v2 with the same network and weights that is there in this repository. I am not able to get the loss reduced much, otherwise it showing good IOU(0.7) and Recall(0.96). Still working on it, haven't verified the result till now.

linzhi123 commented 5 years ago

Can you share your code? @danishansari

danishansari commented 5 years ago

I used YOLO-v2 itself, did not modified any of the code. I am just training for 36 classes instead of 35. So only config file has to be modified. So you can directly use Darknet for training and inference.

linzhi123 commented 5 years ago

That's fine. Thank you!

danishansari commented 5 years ago

@linzhi123 please do share your experience and results(accuracy/map) after training.

linzhi123 commented 5 years ago

ok @danishansari

sergiomsilva commented 5 years ago

I see you are interested in the training script, and I am sorry for not being able to share it with you guys at this moment because I'm really out of time. However I can give you some hints. The only modification we did in the original YOLO training algorithm was considering its data augmentation process. Basically, it reduces too much the image size, which is undesirable in this case, so we controlled it by performing small changes in the source code. We also included synthetically created data, as mentioned in the paper, but you may also achieve nice results without them.

linzhi123 commented 5 years ago

Thanks very much @sergiomsilva

danishansari commented 5 years ago

@sergiomsilva then those modification about augmentation and controlling size reduction must be there in the darknet code in your repository?

sergiomsilva commented 5 years ago

No, the modification is not in our repo.

Tomhouxin commented 5 years ago

32/5000
I want to ask how to use OCR training license plate data alone, I tried, but not so good @linzhi123

linzhi123 commented 5 years ago

Sorry,I'm also trying. @Tomhouxin

Tomhouxin commented 5 years ago

ok，I use tesseract4.0，and you？

xinyuegtxy commented 5 years ago

why not release the ocr training ????? em..............

xinyuegtxy commented 5 years ago

@linzhi123 hello, how do you train the ocr? And how do you prepare the dataset ?

wanglaotou commented 5 years ago

I want to know how to train ocr too. and could we change the code for two layer plate recognition?

PhilipsKoshy commented 5 years ago

ok，I use tesseract4.0，and you？

could you share the details of how you used tesseract4.0? When I tried I get the error saying DPI information not present in the image.

suvarnak commented 4 years ago

I want to know how to train ocr too. and could we change the code for two layer plate recognition?

yes. I wish to train it for plate with 2 rows