Global suppression heuristic: fast GraphCut in GPU for image stitching.

GraphCut algorithm has shown its effectiveness when solving many computer vision tasks. However, its heavy computational nature makes it hard to apply in real-world applications. Many attempts have been made to accelerate GraphCut algorithm, most successfully seen in methods that utilize parallel computing platforms like CUDA. In this paper, we introduce a parallel implementation of push relabel algorithm for GraphCut on CUDA designed for the image stitching problem. Furthermore, we propose global suppression heuristic to boost the convergence process of the algorithm. Experiment results on sets of thermal infrared and RGB images show that our method can be up to 3 times faster than the fastest sequential algorithm while obtaining satisfactory stitched images. Our source code will be soon available for further research. [ABSTRACT FROM AUTHOR]