Closed capilano closed 5 years ago
I think OpenFace is based on dlib's 68 point facial landmarks.
Ah! Forgot I left a question here. I think OpenFace uses dlib for the bounding boxes and they have their own landmark detector because they cite a few papers on their landing page at Github specifically pertaining to facial landmarks .
@capilano you're right. while OpenFace has a dlib dependency the method for landmark detection is quite different. I believe it is significantly more accurate than dlib (I have conducted a number of experiments, but don't have numbers at hand comparing to dlib directly, for most recent results see the OpenFace 2.0 paper - https://www.dropbox.com/s/ki1inrjcjbwm24j/FG2018_camera_ready.pdf?dl=0). While dlib landmark detector is good, it relies on a somewhat older method which has been superceded with more modern landmark detection approaches.
The use of dlib is for face detection and HOG extraction rather than for landmark detection.
@TadasBaltrusaitis Thanks. I have implementing a CNN based facial detector for my application, which is far superior to dlib (in terms of accuracy).
Do you have any comparison metrics vs the 68 point dlib facial landmarks? Thanks