On Compositional Image Alignment, with an Application to Active Appearance Models

Efficient and accurate fitting of Active AppearanceModels (AAM) is a key requirement for many applications.The most efficient fitting algorithm today is Inverse CompositionalImage Alignment (ICIA). While ICIA is extremelyfast, it is also known to have a small convergence radius.Convergence is especially bad when training and testingimages differ strongly, as in multi-person AAMs. We describe“forward” compositional image alignment in a consistentframework which also incorporates methods previouslytermed “inverse” compositional, and use it to developtwo novel fitting methods. The first method, CompositionalGradient Descent (CoDe), is approximately fourtimes slower than ICIA, while having a convergence radiuswhich is even larger than that achievable by direct Quasi-Newton descent. An intermediate convergence range withthe same speed as ICIA is achieved by LinCoDe, the secondnew method. The success rate of the novel methods is 10 to20 times higher than that of the original ICIA method.