Conclusions

We addressed an important issue for the problem of non-rigid registration. We proposed the Feature-Driven framework, relaxing the groupwise requirement for using efficient compositional algorithms such Inverse Compositional and Learning-Based algorithms. We also explained in details the Feature-Driven parameterization for the TPS and the FFD warps. Experiments show that Feature-Driven algorithms are more efficient compared to classical ones with additive update of the parameters. Overall, the best algorithm is the combination of the Feature-Driven framework, the Forward Compositional update of the parameters and the local registration based on Learning. The proposed algorithms make foreseeable accurate real-time surface registration.

Acknowledgements.

We would like to thank Selim Benhimane for his useful advice, in particular for proposing the Gaussian Mixture Model in order to select the weights of the interaction matrices in the piecewise linear model used in the learning-based local registration.