Combining object-based local and global feature statistics for salient object search

What makes a target object stand out and get attended to immediately? Previous work suggests that a salient object carries discriminating features, which make it distinct from its neighborhood. Many models have approached this problem using different methods of distinctness computation. A class of models exploit global feature statistics and adapt a region-based approach to identify distinct patterns, colors and other features in the image. Another category of models make use of local feature statistics and targets pixel/patch based strategies. We propose an algorithm that integrates local and global information by combining region and patch based operation. Our framework enables the extraction of object based local and global feature statistics that makes the salient object stand out from its neighborhood. We present quantitative and qualitative evaluation to show that our method outperforms six state-of-the-art models on traffic signs and object databases.