Can you please provide some advice for the realnets training?

[raduprv]

Hello,
Other than trial and error (which takes a lot of time, being cpu driven), is there any advice on the dataset?
For example, let's say I want to train it to find people.
Is it best to provide only images with one person? Or photos of crowds are ok too?
The people should be in all kinds of positions (sitting, standing, sleeping, eating, etc.)? Is it ok to provide photos with people in the background, or should they primarily be in the foreground? Is it ok if in some samples people are partially out of frame?
What is a good number of photos for positive and background samples?
Any other tips? Perhaps a tutorial too?
Thank you for the great work!