Alien FaceHugger v Predator - Face Tracking Hots Up
Written by Mike James
Sunday, 11 November 2012
A simple video camera as input seems to be getting good enough to challenge depth input devices such as the Kinect. Now we have two competing face tracking methods: the new, aptly named, FaceHugger and the award winning Predator. Can we resist saying it looks like a "face- off".
A few months ago the TLD algorithm created a lot of interest. Developed by Zdenek Kalal as part of his PhD, it took an existing tracking method - Lucas-Kanade - which works well for short term tracking and added some online correction that keeps the tracker more firmly on its target. The Lucas-Kanade tracker works using the equations of optical-flow. These relate the time derivatives between frames to the space derivatives within frames which then allows you to solve the equations to find where a set of pixel values has moved to, i.e the displacement of the target.
All you have to do is draw a bounding box around the object you want to track and the algorithm keeps the bounding box on the target as it moves. What is impressive is not just the way the bounding box sticks to the target but how well it deals with reacquisition and the partial obscuring of the target.
The TLD algorithm is open source and can be downloaded and used, but the research group also has a commercial version, which is called Predator.
Now there is a challenger to Predator, which too has references to sci-fi movies built into its name. The Alien tracker, designed by Federico Pernici MICC - University Of Florence, Italy, is similar to the TLD algorithm but uses Scale Invariant Feature Transforms (SIFT). The idea is to use a lot of local features to identify the object as it moves. A set of features is used for the tracked area and a set for an annular region around the area to provide context - however, the exact way it all works hasn't been published.
What has been published is a comparison of the Alien method against other well-known methods including TLD. And the results:
"ALIEN achieved the best score in the sequences and matched the performance of the current state of the art method"
You can also download and try out the Alien software, but it is only available as an .exe that will run under Windows 7 64 bit. Whether or not we will see the source code or the exact details of how it all works, only time will tell.
Both TLD and Alien can be adapted to track specific objects. When Alien is used for and face tracking, this is referred to as FaceHugger. You can see an example fo FaceHugger in action against Predator in this short video:
If you want to see face tracking demonstrated with the potential for lots of false positives, that's what the next video shows:
Being able to try the method out with a prepackaged executable is no substitute for being able to see the code. Let's hope for more details of how it all works very soon.
The Project Oxford team has released a new demo that uses its machine learning Face detection API. This one recognizes emotional states in photographs of people and seems to be impressively sensitive. [ ... ]