Reference-free monocular 3D tacking of deformable surfaces

We address the problem of deformable surfaces tracking in a monocular video for the purpose of reconstructing the 3D shape of the intended surface. The proposed method is based on structure from motion. In this method, deformation of the surface is modeled as a Gaussian process, and hence the tracking problem is formulated using the predictive capability of stochastic processes. Based on our experimental results, the proposed method is even more accurate than template-based methods. Also, it is highly robust against the tracking initialization, which is a challenging problem in all tracking algorithms.