Automatic Eigentemplate Learning for Sparse Template Tracker

Automatic eigentemplate learning is discussed for a sparse template tracker. It is known that a sparse template tracker can effectively track a moving target using an eigentemplate when it is appropriately prepared for a motion class or for an illumination class. However, it has not been easy to prepare an eigentemplate automatically for any image sequences. This paper provides a feasible solution to this problem in the framework of sparse template tracking. In the learning phase, the sparse template tracker adaptively tracks a target object in a given image sequence when the first template is provided in the first image. By selecting a small number of representative and effective images, we can make up an eigentemplate by the principal component analysis. Once the eigentemplate learning is accomplished, the sparse template tracker can work with the eigentemplate instead of an adaptive template. Since the sparse eigentemplate tracker doesn't require any adaptive tracking, it can work more efficiently and effectively for image sequences in the class of learned appearance changes. Experimental results are provided for real-time face tracking when eigentemplates are learned for pose changes and for illumination changes, respectively.