Improving random walker segmentation using a nonlocal bipartite graph

The classical random walk segmentation explores merely local affinity among neighboring pixels for cutting out objects, which falls short of effectiveness when handling distant repetitive patterns. Meanwhile, the running efficiency is also limited by solving a large scale linear system. To alleviate the quandary, in this paper, we first propose to introduce nonlocal affinity among distant pixels with similar local features in the underlying segmentation graph, which enables label propagation among disconnected foregrounds and thus multiple repetitive patterns can be segmented jointly. Secondly, the segmentation graph is extended to a multi-scale superpixel based bipartite graph, where random walker segmentation is performed to transfer labeling information from annotated seeds to superpixels and further to unlabeled pixels. We show such a bipartite graph based approach can considerably save the computational cost without sacrificing the segmentation accuracy. Extensive experiments are conducted on multiple datasets, demonstrating the effectiveness of the proposed method