alpr-unconstrained problems about ocr dataset

Hello, I'm now trying to train the ocr model using my dataset, but the trained model just doesn't work. I want to know if my annotations are right. Every LP image has several bounding boxes. There's an example:

class_id x y w h 33 0.0970149253731 0.443069306931 0.0985074626866 0.509900990099 16 0.222388059701 0.457920792079 0.10447761194 0.490099009901 9 0.361194029851 0.457920792079 0.113432835821 0.490099009901 32 0.597014925373 0.487623762376 0.107462686567 0.519801980198 2 0.741791044776 0.502475247525 0.110447761194 0.509900990099 16 0.877611940299 0.512376237624 0.101492537313 0.509900990099 I just use darknet to train your ocr network, so could you give me some hints? Plus, can you tell me how many training LP images you used for ocr network?

Dec 07 '18 02:12 xiaogangjia

You need to put the txt file and the picture in the same directory for training. @heuDavidJia

Dec 07 '18 08:12 linzhi123

Take a look at #19 , even with correct labels you will not be able to train the network with Darknet as is. You need to modify the data augmentation part to be less aggressive.

Dec 07 '18 12:12 sergiomsilva

@sergiomsilva Thanks for replying. I will try to modify the data augmentation part.

Dec 08 '18 01:12 xiaogangjia

@linzhi123 I guess the directory of dataset is right, but there are some problems with my data. Have you trained the model successfully?

Dec 08 '18 01:12 xiaogangjia

What do you think is wrong with the data? It's okay for me to use the author's network to train the model, and I only used 1,000 pictures to train, and the effect is good.

Dec 08 '18 04:12 linzhi123

@linzhi123 I guess i have find the problem, the annotation tool i used has a big bug. Do you train the network with Darknet directly or with modified data augmentation part?

Dec 08 '18 07:12 xiaogangjia

Direct use, no modification

Dec 10 '18 02:12 linzhi123

@linzhi123 Have you changed the anchors for your own data? And should i resize my data to 240*80 before training?Cause i always think the darknet has completed the resize work.

Dec 10 '18 11:12 xiaogangjia

@linzhi123 are you using actual images or synthetically generated? I started training it on synthetic data, the accuracy on val and test data is about 99% but on real data it's not not giving good accuracy, and it's not even robust on one frame it's giving different output and on another it's different of the same plate. while after 10k iterations IOU is about 0.9, class score id 0.99, objectness is 0.9 and recall 1.0. Any hint and suggestion will be appreciated.