Face chip detectors/croppers

Before performing the face embedding step required for facial recognition, we need to extract rotated, non-aspect ratio preserving scaled images of an exact square size, since this is the type of data the embedders are trained on.

There are currently two major approaches to extracting face chips: those based on dlib's pipeline and those based on OpenPose skeletons. Typically the latter is recommended if you are planning on extracting these skeleton keypoints at any point, since it is faster and allows the faces to be associated with tracked skeletons.

dlib

The dlib based face chip extractor first runs a face detector and then as show in the pipelines overview

The following modes are available:

Mode name	Face detector	Face keypoint model
dlib-hog-face5	HOG	5-point
dlib-hog-face68	HOG	68-point
dlib-cnn-face5	CNN	5-point
dlib-cnn-face68	CNN	68-point

The CNN detector is slower than the HOG detector, but can be GPU accelerated. The 5 point face keypoint model is slightly faster than the 68 keypoint model. Both keypoint models run on the CPU only.

OpenPose

The OpenPose based face chip extractor uses keypoints detected by OpenPose. If your keypoint model (see meet the pose estimators for a list of keypoint models includes the face, you can use the 68 keypoint model, otherwise you can use the 3 keypoint model which uses only the face keypoints from BODY_25.

Mode name	Face keypoint model
openpose-face3	3-point
openpose-face68	68-point