Strategy to train - questions

1. Do you train for all circumstances (side/profile look, glasses, facial hair, etc)
2. How many images ?
3. Equal amount of images per person or does it not matter?
4. Do you feel if you cater to much to specific items (glasses, jewelry, shots of different lighting, long hair, short hair, beard no beard, young, old) you dilute the model and basically want every model to be the same as the next one.

My take on it currently is:
it is up to you to define the model. So i guess the more specifics you add that is specific to that person, the better the recognition. As long as you keep the images high-res, as otherwise the model will become diluted with image artefacts/noise etc. Once you start adding pics from your RTSP camera it will look recognise very well but it will also say the mailman is the same as family member x.

I've switch to sub-arcnet custom build (non-gpu) but i noticed one difference is that it marks almost everything as unknown, but it also doesn't seem to detect/recognize faces from afar anymore, whereas with the normal build it does.