Graph theory has been proved effective for salient object detection.Existing graph-based salient object detection methods propagate saliency either via a one-stage process,where both seed nodes and graph edges are fixed,or via a two-stage scoring,where seed nodes are updated but the graph edges are still static.The fixed graph edges usually impose a spatial consistency within a small local neighborhood,namely adjacently spatial consistency.Even though this local consistency may be effective in the simple cases but may be ineffective in the nonhomogeneous regions or complex scenes.Therefore,the performance of existing graph-based salient object detectors are in general unsatisfactory when dealing with real-world applications.In this paper,to tackle this problem,we propose two-stage graphs for salient object detection.At the first stage,a regular graph is constructed considering the adjacently spatial consistency,such that the spatial consistency within a local neighborhood can be preserved.On top of it,the coarse detection results obtained at this stage can further locate discriminatively potential foreground candidates and also potential background candidates.At the second stage,a novel graph is constructed by further imposing a new spatial consistency within the foreground candidates and within background candidates,respectively,namely regionally spatial consistency.In another word,we consider not only the adjacently spatial consistency but also the regionally consistency at the second stage.Furthermore,the coarse detection results at the first stage play a vital role and determine the performance of the second stage,which is achieved by a proposed weighted joint robust sparse representation.Particularly,the second stage is generic enough to be integrated in any of salient object detectors,enabling to improve the performance.Finally,experiments on benchmark datasets validate the effectiveness and superiority of the proposed method over the state-of-the-art methods.