Face landmark extraction in OpenCV 3.0. Can anyone suggest any good open source libraries that will allow me to extract facial landmarks?

EiiNHERJii1991 picture EiiNHERJii1991 · Apr 15, 2015 · Viewed 9.6k times · Source

I am currently using OpenCV3.0 with the hope i will be able to create a program that does 3 things. First, finds faces within a live video feed. Secondly, extracts the locations of facial landmarks using ASM or AAM. Finally, uses a SVM to classify the facial expression on the persons face in the video.

I have done a fair amount of research into this but can't find anywhere the most suitable open source AAM or ASM library to complete this function. Also if possible I would like to be able to train the AAM or ASM to extract the specific face landmarks i require. For example, all the numbered points in the picture linked below: www.imgur.com/XnbCZXf

If there are any alternatives to what i have suggested to get the required functionality then feel free to suggest them to me.

Thanks in advance for any answers, all advice is welcome to help me along with this project.

Answer

ekim picture ekim · Jul 9, 2015

In the comments, I see that you are opting to train your own face landmark detector using the dlib library. You had a few questions regarding what training set dlib used to generate their provided "shape_predictor_68_face_landmarks.dat" model.

Some pointers:

  • The author (Davis King) stated that he used the annotated images from the iBUG 300-W dataset. This dataset has a total of 11,167 images annotated with the 68-point convention. As a standard trick, he also mirrors each image to effectively double the training set size, ie 11,167*2=22334 images. Here's a link to the dataset: http://ibug.doc.ic.ac.uk/resources/facial-point-annotations/
    • Note: the iBUG 300-W dataset includes two datasets that are not freely/publicly available: XM2VTS, and FRGCv2. Unfortunately, these images make up a majority of the ibug 300-W (7310 images, or 65.5%).
    • The original paper only trained on the HELEN, AFW, and LFPW datasets. So, you ought to be able to generate a reasonably-good model on only the publicly-available images (HELEN,LFPW,AFW,IBUG), ie 3857 images.
      • If you Google "one millisecond face alignment kazemi", the paper (and project page) will be the top hits.

You can read more about the details of the training procedure by reading the comments section of this dlib blog post. In particular, he briefly discusses the parameters he chose for training: http://blog.dlib.net/2014/08/real-time-face-pose-estimation.html

With the size of the training set in mind (thousands of images), I don't think you will get acceptable results with just a handful of images. Fortunately, there are many publicly available face datasets out there, including the dataset linked above :)

Hope that helps!