Object extraction is a core component of computer vision with application to segmentation, tracking, etc. In this paper we propose a GPU graph-based approach to object segmentation. The main contributions of our approach consist of an adaptive, evolving schema to update to statistical properties of the object and the use of a local variant push-relabel algorithm to recover the global minimum of the designed cost function. Furthermore, we propose the implementation of the method on a graphics processing unit where each node of the graph is considered as an independent processor that is connected with the neighborhood nodes. Such a schema recovers segmentation in an optimal and progressive manner. Promising experimental results and important decrease on the computational complexity demonstrate the potentials of our approach.

Abstract not found