First, we compute the image saliency using and construct an importance map based on it. The image saliency by was developed to measure visual attractiveness and used for human detection, but not designed to be consistent with the image structure, e.g. the pixels from the same object can have very different saliency. Therefore we define an importance map that rep- resents the saliency as well as has the consistent structure with the original image, i.e. the pixels on the same object should be likely assigned the similar importance.