Most of the saliency methods are evaluated on their ability to generate\nsaliency maps, and not on their functionality in a complete vision pipeline,\nlike for instance, image classification. In the current paper, we propose an\napproach which does not require explicit saliency maps to improve image\nclassification, but they are learned implicitely, during the training of an\nend-to-end image classification task. We show that our approach obtains similar\nresults as the case when the saliency maps are provided explicitely. Combining\nRGB data with saliency maps represents a significant advantage for object\nrecognition, especially for the case when training data is limited. We validate\nour method on several datasets for fine-grained classification tasks (Flowers,\nBirds and Cars). In addition, we show that our saliency estimation method,\nwhich is trained without any saliency groundtruth data, obtains competitive\nresults on real image saliency benchmark (Toronto), and outperforms deep\nsaliency models with synthetic images (SID4VAM).\n