Abstract.
Active Appearance Models (AAM) are compact represen- tations of the shape and appearance of ob jects. Fitting AAMs to im- ages is a diffcult, non-linear optimization task. Traditional approaches minimize the L2 norm error between the model instance and the input image warped onto the model coordinate frame. While this works well for high resolution data, the fitting accuracy degrades quickly at lower resolutions. In this paper, we show that a careful design of the ?tting criterion can overcome many of the low resolution challenges. In our resolution-aware formulation (RAF), we explicitly account for the finite size sensing elements of digital cameras, and simultaneously model the processes of ob ject appearance variation, geometric deformation, and im- age formation. As such, our Gauss-Newton gradient descent algorithm not only synthesizes model instances as a function of estimated parame- ters, but also simulates the formation of low resolution images in a dig- ital camera. We compare the RAF algorithm against a state-of-the-art tracker across a variety of resolution and model complexity levels. Ex- perimental results show that RAF considerably improves the estimation accuracy of both shape and appearance parameters when ?tting to low resolution data.