facenet-pytorch VS CompreFace

It's generally better to split the task into a multiple tasks. First I'd want to detect and extract faces. There are a number of pretrained models that you could use for that, e.g. https://github.com/timesler/facenet-pytorch, https://github.com/opencv/opencv/tree/master/data/haarcascades. Once you've extracted faces, you can train a facial recognition using something like a siamese network as you normally would.

For face recognition - CompreFace. Disclaimer - I created it, as an alternative you can use MachineBox, but it's not open source and has limits. Also, I think, you will use some software to control the system, e.g. Frigate or Home Assistant, I think this repository can be useful for you.

If you have a really simple application, you can just save the encodings into the files. If not - it's better to use a database. SQL is ok. But for the best results, I would suggest using milvus.io, as it was created for saving vectors and finding the distances (I haven't tried it, though). If your final goal is not to learn face recognition basics, you can just use free ready to use solutions like CompreFace and Machinebox

Free and Open-Source Face Recognition System that can be integrated into any system without prior AI knowledge: https://exadel.com/solutions/compreface/

Have you considered CompreFace?

Also found out about the the[CompreFace](https://github.com/exadel-inc/CompreFace) project and added support for their facial recognition. So far the results have been pretty promising!