clmtrackr

clmtrackr is a javascript library for fitting facial models to faces in videos or images. It currently is an implementation of constrained local models fitted by regularized landmark mean-shift, as described in Jason M. Saragih's paper. clmtrackr tracks a face and outputs the coordinate positions of the face model as an array, following the numbering of the model below:

Reference - Overview

The library provides some generic face models that were trained on the MUCT database and some additional self-annotated images. Check out clmtools for building your own models.

The library requires jsfeat.js (for initial face detection) and numeric.js (for matrix math).

For tracking in video, it is recommended to use a browser with WebGL support, though the library should work on any modern browser.

For some more information about Constrained Local Models, take a look at Xiaoguang Yan's excellent tutorial, which was of great help in implementing this library.

Examples

• Tracking in image
• Tracking in video
• Face substitution
• Face masking
• Realtime face deformation
• Emotion detection
• Caricature

Usage

Download the minified library clmtrackr.js and one of the models, and include them in your webpage. clmtrackr depends on numeric.js and jsfeat.js, but these are included in the minified library.

/* clmtrackr libraries */
<script src="js/clmtrackr.js"></script>
<script src="js/model_pca_20_svm.js"></script>

The following code initiates the clmtrackr with the model we included, and starts the tracker running on a video element.

<video id="inputVideo" width="400" height="300" autoplay loop>
  <source src="./media/somevideo.ogv" type="video/ogg"/>
</video>
<script type="text/javascript">
  var videoInput = document.getElementById('inputVideo');
  
  var ctracker = new clm.tracker();
  ctracker.init(pModel);
  ctracker.start(videoInput);
</script>

You can now get the positions of the tracked facial features as an array via getCurrentPosition():

<script type="text/javascript">
  function positionLoop() {
    requestAnimationFrame(positionLoop);
    var positions = ctracker.getCurrentPosition();
    // positions = [[x_0, y_0], [x_1,y_1], ... ]
    // do something with the positions ...
  }
  positionLoop();
</script>

You can also use the built in function draw() to draw the tracked facial model on a canvas :

<canvas id="drawCanvas" width="400" height="300"></canvas>
<script type="text/javascript">
  var canvasInput = document.getElementById('drawCanvas');
  var cc = canvasInput.getContext('2d');
  function drawLoop() {
    requestAnimationFrame(drawLoop);
    cc.clearRect(0, 0, canvasInput.width, canvasInput.height);
    ctracker.draw(canvasInput);
  }
  drawLoop();
</script>

See the complete example here.

Development

First, install node.js with npm, then install Grunt.

In the root directory of clmtrackr, run npm install then run grunt. This will create clmtrackr.min.js and clmtrackr.js.

To test the examples locally, you need to run a local server. One easy way to do this is to install http-server, a small node.js utility: npm install -g http-server. Then run http-server in the root of clmtrackr and go to http://localhost:8080/examples in your browser.

License

clmtrackr is distributed under the MIT License