search

bbonik / facial-landmark-frontalization

Function to frontalize non-frontal 2D facial landmarks generated from the DLIB library
MIT License
21 stars 10 forks source link

Dlib #4

Closed jayp84847 closed 3 years ago

jayp84847 commented 3 years ago

Yet I am using dlib 68_face_landmarks.dat in my project but it's not stable and accurate so any suggestions to make it more stable and accurate.

bbonik commented 3 years ago

Can you elaborate more on what do you mean by "stable" and "accurate"? You refer to the initial DLIB landmarks or the frontalized ones? From my experience, DLIB landmarks are not stable. They fluctuate a lot and are affected by uneven illumination (like shadows or directional light), facial hair and dark skin. This has to do probably with the dataset that they were trained with, which, as users, we cannot control. The best conditions for DLIB landmarks to be stable are, black facial features over white-tone skin face, no facial hair or accessories, illumination behind the camera, no illumination behind or on top/sideways of the face. Any deviation from these conditions makes the DLIB landmarks fluctuate. The frontalized landmarks are based on the original DLIB landmarks. So, if the original landmarks are fluctuating a lot, then these fluctuations will also affect the frontalized landmarks.

jayp84847 commented 3 years ago

Stable and accurate mean by I want to make as stable as snapchat facial landmarks. Like dlib's 68_face_landmarks.dat module works, but not working for rotational faces and even not working when user tries with glasses.

Also cause of less accuracy, when I deployed face mesh of cat(example) it seems like vibrating. Not stable.

Can you share your mail. I will share you my current result video. Unable to send it on github.

bbonik commented 3 years ago

You can contact me at bbonik@gmail.com