Are pixelated frigate face pictures okay for training?

[mtthidoteu]

Hello!
I have a question regarding the training.
I have uploaded 5 photos for each person taken from my Photo Gallery and I am incredibly amazed by how good the performance is. It picks up blurry pixelated versions of the face and recognises it as the person. There are few cases where it sits at 70%, and the docs say to use these to train. However, they are very blurry and pixelated.

If I go into the Face recognition tab, I see all the faces and as they are from my cameras, the quality is not amazing and all my cameras's face pictures end up being a blurry zoomed. For Double Take, the advice was only to use high quality photos for the training.

In frigate, should I use these pixelated photos taken from the cameras with the assumption that "You train what you want to see"?

Thanks for your advice!

[hawkeye217]

See the official documentation: https://deploy-preview-16390--frigate-docs.netlify.app/configuration/face_recognition/#creating-a-robust-training-set

[mtthidoteu]

That's the thing, I did look at the documentation but I feel like every single "face" snapshot, even on a high res camera would be very pixelated vs. a nice photo taken from a camera/phone, so what constitutes a good photo for training?

Should we only ever train from pictures that are not from the cameras?

[NickM-27]

The main point is that you want to be able to make out the details of a persons face. So as long as it is clear where the eyes, nose, and mouth are (not smudges) then it is a fine photo. Like it says in the docs only 1-5 phone camera images are recommended.

[hawkeye217]

Should we only ever train from pictures that are not from the cameras?

Yes. As the documentation states, you should start with a small number (1-5) of higher resolution photos per person. See the steps here: https://deploy-preview-16390--frigate-docs.netlify.app/configuration/face_recognition/#step-1---building-a-strong-foundation