Can classification models replace LPRnet

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) RTX3050
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) Classification, LPRnet
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
• Training spec file(If have, please share here)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

Hi, i have some questions to ask about replacing the LPRnet with Classification models. Just wanna ask can i train a classification model to detect license plate instead of using LPRnet.

Lets just say if i only want to detect 10 different license plate only, where each license plate have a different FOV in my training set, and each license plate contains 50 different images. In my label file, i will only put 10 labels for this license plate. For example, “ABC123”, “EFG567”, etc. Is it recommended and what are the drawbacks of using a classification model instead of a LPR model. Besides, what should i take note if i use a classification model to use as a license plate recognition model. Please advice.

2

If there are only 10 different license plates, for example, “ABC123”, “EFG567”, “LMN489” , etc. Yes, a classification model can classify them.
The training dataset should contain 10 folders:

ABC123
       |--- 1.jpg (contains “ABC123” license plate)
       |--- 2.jpg (contains “ABC123” license plate)
       |--- 3.jpg (contains “ABC123” license plate)
       |--- ...
 
EFG567
       |--- 1.jpg (contains “EFG567” license plate)
       |--- 2.jpg (contains “EFG567” license plate)
       |--- 3.jpg (contains “EFG567” license plate)
       |--- ...
 
LMN489
       |--- 1.jpg (contains “LMN489” license plate)
       |--- 2.jpg (contains “LMN489” license plate)
       |--- 3.jpg (contains “LMN489” license plate)
       |--- ...
...

The drawback is that the classification model can only classify these 10 classes of license plates.

3

Understand. May i know what is the suggested training set size for each classes. Lets say this 10 classes of license plates, i may have 15 images for each classes. What is i need to train for 1 million classes, is 15 images enough for each class?

4

You can refer to ILSVRC2012 dataset. The dataset has 1000 classes, and each class has about 1200 images.
So for your case, it is a bit less. You can generate more by using cv method. Or even you can copy them in order to do oversampling.

5

Lets say for the license plate recognition, what are the drawbacks of using classification methods for 10million of license plate combinations. If classification is used for such an amount of classes, will the accuracy drop? and will the inference performance be slower

6

Actually we did not do this kind of experiment to run classification training against lots of license plates dataset. I think the accuracy is not good comparing to LPRNet. LPRNet can get more than 99%. Classification model for top1 or top5 may be good. But if the N is larger, the topN’s accuracy will not be good. Also, it is not convenient for using classification model to do LPR because a new license plate cannot be classified.
For inference fps, you can compare the LPR model and classification model from the info mentioned in model card.

7

Thanks for the output given

8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.