During the model learning phase, each image is over-segmented into several regions. Then, the patch-level feature, region-level feature, mean saliency value and location index are used as the input of GMS to learn the model parameters. The features of the extracted objects are used to train SVM classifier