Deformable Template Tracking in 1ms

The goal of template tracking is to estimate the transformation parameters that define the motion of the planar template. Typically, the transformation parameters encode linear transformation based on homographies with 8 degrees of freedom, which allows them to track rigid motions of the template. However, when it comes to tracking surfaces that undergo non-rigid deformations, deformable transformations with higher degrees of freedom must be used. Most works on deformable template tracking rely on feature points [1, 5, 6, 7, 8]. It follows the traditional algorithm where it detects feature points on the current frame; finds feature point correspondences between the current frame and the template; remove outliers from these correspondences; and, estimate the deformation. However, since deformable models have a higher degrees of freedom, it becomes more difficult to detect outliers. As a result, it requires a longer runtime. Therefore, this paper aims to address the problem of real-time deformable template tracking. Instead of using tracking-by-detection approaches with feature points, our work focuses on a frame-to-frame tracking approach with a dense pixel arrangement on the template. Hence, to track the template, we use the linear predictor A which establishes a linear relation between the vector of image intensity differences δ i of a template and the corresponding template transformation parameters δ μ , which is written as [3]: