Does l2 distance reliable for comparision?

[Alexander Ustimenko]

Hi!

We use these models/params for face recognition:

• facePredictor = shape_predictor_68_face_landmarks.dat
  
• model = nn4.small2.v1.t7
  
• imgDim = 96 

Each representation is array of 128 length.

At demo page for comparision: https://github.com/cmusatyalab/openface/blob/7659f7907d1d1aa16ba11423500a6d927106b742/demos/compare.py#L100:L104

we see l2 distance as a similarity measure.

We have found, that more or less reliable similar faces are below 0.16 distance.

Also we found that two different photos of same person has 0.59 distance.

So my questions are:

1. Does l2 distance between normalized 128-dimensional unit hyperspheres reliable for comparision? Or at least for sorting by similarity.

2. In our important-for-us case where we see distance 0.59 -- what could be the reason of such distance? Pose? Something other? Or 128-dimensional unit hypersphere already normalized for all poses/pitches/etc?

Tnx!