Deformable template tracking in 1MS

We address the problem of real-time deformable template tracking. Our approach relies on linear predictors which establish a linear relation between the image intensity differences of a template and the corresponding template transformation parameters. Up to this work, linear predictors have only been used to handle linear transformations such as homographies to track planar surfaces. In this paper, we introduce a method to learn non-linear template transformations that allows us to track surfaces that undergo nonrigid deformations. These deformations are mathematically modelled using 2D Free Form Deformations. Moreover, the simplicity of our approach allows us to track deformable surfaces at extremely high speed of approximately 1 ms per frame that has never been shown before. To evaluate our algorithm, we perform an extensive analysis of our method's performance on synthetic and real sequences with different types of surface deformations. In addition, we compare our results from the real sequences to the feature-based tracking-by-detection method [20], and show that the tracking precisions are similar but our method performs 100 times faster.