Why does the Windows ANPR SDK have much lower accuracy compared to your online recognition service?

Hair-S愛爾絲

unread,

Oct 21, 2025, 1:16:03 AMOct 21

to doubango-ai

Hello Sir,

I intend to purchase the ANPR Runtime License, so I downloaded the Windows version SDK from GitHub for testing.
However, I noticed that when using the same license plate images, the recognition accuracy of the SDK is significantly lower than that of your official online recognition service.

Could you please confirm whether this difference is because the GitHub version is a trial or limited feature version?

Thank you for your assistance.
Best regards,
James Ou

Mamadou DIOP

unread,

Oct 21, 2025, 1:22:27 AMOct 21

to Hair-S愛爾絲, doubango-ai

Hello,

"the recognition accuracy of the SDK is significantly lower than that of your official online recognition service." -> this affirmation is not true at all and has been discussed here several times.

Provide here significant number of images to back your claim if you want us to investigate.

The online demo and the SDK use the same models and have same accuracy. The returned results may be different on some images for various reasons. One reason: the online demo use TensorRT for inference on GPU while you're using OpenVINO or Tensorflow on CPU. Another reason: Linux versus Windows. Etc etc

Sent from my Galaxy

--
You received this message because you are subscribed to the Google Groups "doubango-ai" group.
To unsubscribe from this group and stop receiving emails from it, send an email to doubango-ai...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/doubango-ai/eca1ac33-29af-4b1e-a6ca-cbcc354a97f8n%40googlegroups.com.

Hair-S愛爾絲

unread,

Oct 21, 2025, 3:47:11 AMOct 21

to doubango-ai

Hello Sir,

Thank you for your reply.
According to your explanation, the recognition accuracy between the SDK and the online demo is not significantly different. Therefore, we downloaded the latest version of the SDK from GitHub and changed the “recogn_rectify_enabled” setting from False to True.
After this adjustment, the recognition accuracy improved noticeably, although the CPU load also increased.
Overall, the recognition accuracy of the SDK downloaded from GitHub is quite comparable to that of the online demo.

Thank you for your assistance.

Best regards,
James Ou

Mamadou DIOP 在 2025年10月21日星期二下午1:22:27 [UTC+8] 的信中寫道：