real time inference

[nicholasguimaraes]

Hello @Talgin , firstly thank you for the amazing work done!

Is there any script to do inference in real time or using a .mp4 video as shown in the insightface repository?

[Talgin]

Hi @nicholasguimaraes

You can use https://github.com/Talgin/stream_worker, but it just does detection, so you have to write the recognition code after the detection.
Write me if any help needed, currently I have some big tasks to finish, but I will try to help as much as I can.

[nicholasguimaraes]

Thank you for that! @Talgin
I'll be adapting the code to do real time face recognition. I'll post it here as soon as it is done so you can share with more people.

Another question, what should be the minimal number of images of one person? Two(for the training pair) + images for validation during training?

[Talgin]

Hi @nicholasguimaraes ,
Most people recommend 5+ images per person. In our case we used at least 10 images per person strategy and the outcome was good (98+ accuracy).
Thank you! :)