Texture Synthesis Method for Generic Video Sequences

Abstract

An effective texture synthesis method is presented that is inspired by the work of Kwatra et al. [1]. Their algorithm is non-parametric and patch-based. Blending between overlapping patches is optimized using graph cut techniques. We generalize the initial approach [ 1 ] to achieve a new synthesis algorithm that yields improved results for a much larger class of natural video sequences. For that, two major extensions have been provided: (1) the ability to handle constrained texture synthesis applications and (2) robustness against global camera motion. Constrained synthesis thereby refers to integrating synthetic textures into natural video sequences, as opposed to unconstrained texture synthesis, where (infinite) spatio-temporal extensions of single textures are generated. Camera motion compensation enables applicability of the synthesis algorithm to video sequences with a moving camera. The results presented in this paper show that the proposed improvements yield significant subjective gains compared to the initial algorithm.