Deep learning for image based shelve inventories

Deep learning is applied to accurately detect drugstore products in images of supermarket shelves. Detection here means finding precise and tight bounding boxes for the objects. The algorithm learns to fit these frames step-wise better to the objects. More precisely, the so-called agent can zoom, translate, change the aspect ratio, and divide the bounding box. The reward function guiding the agent through its search is adjusted to these degrees of freedom.