A Two-Stage Approach for AAM Fitting

The active appearance model (AAM) is a powerful generative method for modeling and registering deformable objects. The project-out inverse compositional (POIC) method is one of the most common methods for AAM fitting and is also the fastest known method to date. However, it does not work well when the initialization is far from the optimum. In this paper, a two-stage fitting procedure based on POIC method is presented: at the first stage, we consider the model as a rigid body and roughly fit it to obtain large capture range; at the second stage, we deem the model as a deformable model and solve for local deformations to achieve accurate results. In other words, the algorithm firstly updates the global similarity transform parameters and then updates the global similarity transform parameters and the linear shape parameters simultaneously. Experimental results demonstrate that our method can achieve larger capture range with better convergence rates.