What do you think about AAM's?

Edited by author 03-25-2006 08:22 PM
A really comprehensive introduction to AAMs and extension of the Cootes and Taylor approach can be found in the "Active Appearance Models Revisited" paper by Iain Matthews and Simon Baker(CMU).
http://www.ri.cmu.edu/pub_files/pub4/matth...ews_iain_2004_2.pdf

This also makes an interesting (and fairly short) reading:
http://www.isbe.man.ac.uk/~bim/Papers/asm_aam_overview.pdf

Problems with this approach are: it requires expensive training data with marked landmarks, it requires a good initialization, and it doesn't easily scale to multiple object classes. AAMs are probably very good for face tracking, but I doubt they are powerful enough for 'object recognition.'

As with most model based approaches, AAMs too require training data with hand labeled landmarks and for every object we need to build a separate model to perform recognition. However, AAMs can be very useful for "face" recognition and face expression analysis. Tasks such as "registration" and "pose normalization" become easy with model based approaches.

Edited by author 03-28-2006 09:19 PM
One of the really good theses on Face Recognition Across Pose by Vetters group:
http://gravis.cs.unibas.ch/publications/pami03.pdf
The results are on the CMU PIE(Pose Illumination Expression) dataset and the FERET dataset.

Although it may not work for any object class, maybe it would be useful for specific classes similar to faces, such as cars. Could AAMs be used to easily find intraclass variation within cars and perhaps find them in a scene?

As Tomasz mentioned, a good initialization is required for the use of this method. This confuses me because the last sentence of the abstract and the paper itself states that the intended application of this work is "locating deformable objects". Since AAMs need to rely on some other method for estimating the location of the objects, it seems like they would be better suited for tasks that use the converged model parameters to describe intraclass variation. The person or expression could be identified by matching in parameter space after the best fit is found.

One could imagine building some sort of bag-of-texton-words classifier that can predict whether a face or a car exists inside an image window, and then attempt to fit an AAM model for that object class inside the window (like Nik suggested).

When attempting to use AAMs for cars, it appears that one would need a collection of spatial models (one for each aspect). Faces are somewhat easy, because they are essentially 2D objects that are rich with internal features (the eyes, nose, and mouth are almost always seen in the same spatial configuration) while cars look differently from different points of view.

If one can do what I outline above, then I imagine that AAM-style model fitting can be useful for distinguishing cars from buses or cows from horses (something that bags of words might not be good at). One could probably use one AAM for car/bus like things and another AAM for four legged mammals. Then the distinction between cars and buses or cows and horses could be achieved by building a classifier on top of the AAM parameter space.