Rework face recognition accuracy test

[marcelklehr]

Fetches 20 famous people, fetches 10 pictures of each from google, runs the face detection on them to produce vectors, calculates the centroid for each person and the distance of each vector from the centroid. Also fetches 30 pictures for the query 'portrait' from google to mimic a kind of "normal" distribution of distractor faces and runs face detection to produce vectors and calculates the distance from the person's centroid. Vectors within 0.45 euclidian distance of the centroid are considered positive matches, any vectors further away are negative matches. If true positive rate or true negative rate are below 0.8 / 80% this CI run fails.