Closed Avadesign-David closed 2 years ago
As explained at https://www.doubango.org/pricing.html, the Server (x64) models are more accurate than the Mobile (ARM) models. So, that explains why the result is different from the cloud. You'll have the same result as our cloud if you try on Windows or Linux x86 instead of Raspberry Pi. In your case the issue is that we only have very few license plates from Taiwan. If you have samples (more than 1K) we can add them to the training set to increase the accuracy.
Also consider enabling the rectification layer to increase the accuracy.
Another note: Update your code to 3.3.2: https://github.com/DoubangoTelecom/ultimateALPR-SDK/commit/0b2f9568d58abb9935e1a3fd71c80bc4ba4ce807
Tried with https://www.doubango.org/webapps/alpr/ and only one Q is misinterpreted as 0 and based on the angle it's not really a bug as it's hard even for a human.
I have tried your images with https://platerecognizer.com, I also have 1 error on a different image.
OpenALPR online demo is no longer available but you would probably have the same number of errors.
You should reconsider the position of your camera, enable rectification layer or use a Linux/Windows x64 version.
As explained at https://www.doubango.org/pricing.html, the Server (x64) models are more accurate than the Mobile (ARM) models. So, that explains why the result is different from the cloud. You'll have the same result as our cloud if you try on Windows or Linux x86 instead of Raspberry Pi. In your case the issue is that we only have very few license plates from Taiwan. If you have samples (more than 1K) we can add them to the training set to increase the accuracy.
Also consider enabling the rectification layer to increase the accuracy.
Another note: Update your code to 3.3.2: 0b2f956
Okay, I understand. Although we do not have more than 1K images data, we have about 100 images. And also we can provide the characters(Taiwan license plate used) sample if needed. If these can help with improve accuracy?
As explained at https://www.doubango.org/pricing.html, the Server (x64) models are more accurate than the Mobile (ARM) models. So, that explains why the result is different from the cloud. You'll have the same result as our cloud if you try on Windows or Linux x86 instead of Raspberry Pi. In your case the issue is that we only have very few license plates from Taiwan. If you have samples (more than 1K) we can add them to the training set to increase the accuracy. Also consider enabling the rectification layer to increase the accuracy. Another note: Update your code to 3.3.2: 0b2f956
Okay, I understand. Although we do not have more than 1K images data, we have about 100 images. And also we can provide the characters(Taiwan license plate used) sample if needed. If these can help with improve accuracy?
You can share your images regardless the number. We'll update the model in the coming days.
Okay. Thanks. How can we provide the images?
Google drive or any other sharing method
Okay. We have uploaded the images. It contains about 1k images. Download link: https://avadesign-download.s3.amazonaws.com/Car.zip
Will be fixed in the next version (probably v3.7). Already pushed at https://www.doubango.org/webapps/alpr/
Fixed by v3.7.0
We have encounter accuracy problem with raspberry pi. It recognize the wrong word "Q" -> "0", "Z" -> "7". But some of the image result is correct while we test it on https://www.doubango.org/webapps/alpr/
P.S. Platform: raspberry pi 4 (32bit) SDK version: 3.3