PatchMatch algorithms for motion estimation and stereo reconstruction

Abstract

Correspondence problems are difficult and fundamental topics in computer vision, which aim to find the displacement field between two consecutive images within an image sequence. Both stereo matching and motion estimation belong to the domain of correspondence problems. A common technique is to define a parametric model and estimate its parameters via some optimization algorithms. In this context, this thesis extends the recently proposed PatchMatch algorithm, which is originally designed for finding approximate nearest neighbors, to model parameter estimation. Specifically, there are three purposes of this thesis: (i) We analyze and extend the PatchMatch algorithm to model parameter estimation. (ii) Afterwards, some commonly used parametric models for motion estimation and stereo matching in the literature are reviewed. (iii) Finally, the extended PatchMatch algorithm is implemented and applied to estimate the model parameters summarized above. The estimation performance is evaluated and compared with some other methods in the literature based on the three public benchmarks: Middlebury, KITTI and MPI Sintel.