Automatic learning sparse correspondences for initialising groupwise registration

We seek to automatically establish dense correspondences across groups of images. Existing non-rigid registration methods usually involve local optimisation and thus require accurate initialisation. It is difficult to obtain such initialisation for images of complex structures, especially those with many self-similar parts. In this paper we show that satisfactory initialisation for such images can be found by a parts+geometry model. We use a population based optimisation strategy to select the best parts from a large pool of candidates. The best matches of the optimal model are used to initialise a groupwise registration algorithm, leading to dense, accurate results. We demonstrate the efficacy of the approach on two challenging datasets, and report on a detailed quantitative evaluation of its performance.