I have been studying the gen2 face recognition code and I am still working my way through. I was wondering if someone might know how to change the way the face recognition model is trained? For me, I am not usually able to train the model by capturing a dedicated set of images of someone. I need to be able to record images of anyone who comes within range of the camera and label them later. Gen 2 Face Recognition wants you to create a .npz file for each person you want to recognize. For me, it is better if I have recorded images that I can go back and look at later then label. I am guessing that I would need to capture the images and put them in something like a mp4 file and then when I can, come back read that mp4 file and convert it into the necessary format. This also would require being able to update the "person face file" adding additional images to improve accuracy.

All help is welcomed!

Thanks,
Doug

  • erik replied to this.
    5 days later

    Hi Aeroclass ,
    You could do that as well. Please note that these apps are just experiments/demos that we developed and far from production-ready applications in terms of feature-completeness and robustness/accuracy.
    One example would be to later feed the mp4 (instead of live data) to the pipeline, and vectorized face features would get saved into the DB. You could also later save these vectors if you determine this is a particular person.
    Thoughts?
    Thanks, Erik

      Hi Erik,
      First off thanks for posting your reply I appreciate you taking the time. I completely understand these are NOT production ready application and, as you described just demos etc. That is fine with me as I have a starting point and now, I can expand it etc. That's what I wanted!
      I am sure I can have the app read a mp4 file and feed the frames to the pipeline. That is pretty easy. When you speak of a database, I assume you mean one of the .npz files? Just as you do when creating a vector filled file of a particular person? Obviously, the number of frames of any one person will be pretty small as they are not necessarily doing a pose to build a larger file. Most likely I would get a few seconds of someone in a video and that would happen maybe once or twice a day. So, it will take time to build up enough data for real accuracy. I think that is to be expected.

      Thanks,
      Doug

        Hi Aeroclass ,
        For the demo we just used .npz file, for production you would likely want something more scalable so it can find the person even it if has thousands (or millions) of DB entries - with the current approach that would likely take hours, not milliseconds that would be required for such application. Otherwise yes, that should work.
        Thanks, Erik