Dlib shape predictor models

Dlib [2] is a fantastic C++ library for machine learning, and image processing among others. At Wisimage, we have used it extensively for machine learning and for testing out some fancy algorithms.

One of dlib's most popular features seems to be the shape predictor, the implementation of Kazemi's 2014 CVPR paper [1].

Ensemble of Regression Trees (ERT) is indeed a very efficient and surprisingly precise face landmarks algorithm. Check it for yourself in this video by Vahid Kazemi.

As of 2018, ERT is still a good choice, even against deep learning based algorithms, especially when considering its very low complexity [3].

 Dlib is shipped with the standard iBug 68 points model. However, depending on the application, you might want to have a larger number of points, or at different locations.

You may want to re-train with a different dataset. But in case you don't, Wisimage is making available three models compatible with Dlib:
  • 194 points, HELEN database format. Trained with HELEN + mirrors (183 Mo)

Download the model from here.
  • 75 points, in-house model and data (71 Mo)
 Download the model from here.
  • 77 points MUCT model, Trained with MUCT database, from Milborrow [4] (72 Mo)

Download the model from here.

These models are distributed under the Creative Commons ShareAlike licence (CC-By-SA 2.5). This means that you have to give appropriate credit (this blog post), and share it with this very same licence if you intend to distribute it.

  • [1] One Millisecond Face Alignment with an Ensemble of Regression Trees by Vahid Kazemi and Josephine Sullivan, CVPR 2014 
  • [2] Dlib-ml: A machine learning toolkit, Davis E. King, Journal of Machine Learning Research 10 (Jul), 1755-1758, 2009
  • [3] Face alignment in-the-wild: A survey X Jin, X Tan. Computer Vision and Image Understanding 162, 1-22, 2017
  • [4] The MUCT Landmarked Face Database, S. Milborrow and J. Morkel and F. Nicolls, Pattern Recognition Association of South Africa, 2010


Popular Posts