resize of non-square bounding boxes

[gschramm]

When a detected face bounding box is non-square, the resizing in

facenet-pytorch/models/utils/detect_face.py

Line 312 in 555aa4b

out = cv2.resize(

leads to distortions.
E.g. if the bouding box has width 300:200 (aspect ratio 1.5), and gets resized to 160:160, the face image gets
squeezed in the vertical direction

As a fix, I would suggest to first pad the non-square bounding box to a square size before resizing.

[timesler]

This is deliberate as the pretrained model was trained using images prepared this way