Geodesic Weighted Bayesian Model for Saliency Optimization

Abstract

Bottom-up methods and general Bayesian framework for saliency detection commonly suffer from two drawbacks. First, they are sensitive to background noise, thus background regions similar to objects are also highlighted. Second, they only consider appearance features and thus object with several different parts will not be highlighted uniformly. In this paper, we propose a novel and unified geodesic weighted Bayesian model which considers spatial relationship by reformulating the Bayes' formula. First, we infer a more precise initial salient regions via fully connected CRF model. Second, to highlight the whole object uniformly, we learn a robust measure of region similarity which describes the probability of two regions belonging to the same object, so regions belonging to the same object will be given similar saliency value. Third, using our learnt region similarity as edge weight, we construct an undirected weighted graph to compute geodesic distance of regions. Regions with short geodesic distance from initial salient regions will be attached more importance, thus suppressing background noise. By using results of existing methods as prior distribution, our model can integrate into all methods and improve their performance. Experiments on benchmark datasets demonstrate that our model can significantly improve the quality of saliency detection.

Paper

Pattern Recognition Letters (PDF)
@article{wang2016geodesic,
  title={Geodesic Weighted Bayesian Model for Saliency Optimization},
  author={Wang, Xiang and Ma, Huimin and Chen, Xiaozhi},
  journal={Pattern Recognition Letters},
  volume={75},
  pages={1--8},
  year={2016},
  publisher={Elsevier}
}

Results

We test our method on three typical benchmark datasets: ASD, CSSD and DUT-ORMON. ASD which contains 1000 images is widely used and relatively simple, CSSD contains 200 images which are moderate difficult, DUT-ORMON contains 5168 images which are more challenging.

Dummy Image

Comparison of different methods with their improved versions (*). The first row are tested on ASD, the second are on CSSD, and the third are on DUT-OMRON. The first two columns show the improvement of PR curves, the third column shows the improvement of F-measure, and the last column shows the decrease of mean absolute error (MAE).

Code