Region of interest for Haar samples?

edit

For Viola & Jones algorithm, you can take a look at the corresponding paper and at the dataset they used.

You can also check at some face detection datasets (be careful at the dates):

• Face Databases
• FDDB: Face Detection Data Set and Benchmark
• The Extended Yale Face Database B
• and many more.

Just my opinion as a non expert: you should crop between your two examples as the left example contains too much background in my opinion and the right one seems to me too cropped. As the training process will select relevant features (at least with old algorithms like Viola & Jones, I don't know how it is with DNN algorithms), some litlle background should be harmless I think. The faces have to be consistent, always at the same size and at the same position in the dataset.