Published Paper Details:

Everyday objects or stores by using image based mobile interactions through their devices. For example, as users walk on the street, they might simply point the mobile camera to a store on the street to quickly access its related information, inquire special offers, and make reservations through their mobile devices without physically entering the store. Users depend on mobile device to maintain information and update. OPSs (On Premises Signs) shows great visual diversity accompanied with complex environmental conditions. Consider an example, user walk on the street and he simply point his mobile camera to a store to quickly access its related information, inquire special offers, and make reservations through his mobile without physically entering on that store. Street view scenes are commonly captured by customers devices and they have more real-world characteristics lacking in most existing image datasets, e.g. perspective distortion, foreground and background clutter, etc. To learn a reliable OPS model for recognizing OPSs, a labeled dataset with a huge amount of real-scene images is required. However, precisely labeling OPS categories and regions, i.e., generating strong labels for Learning involves a significant amount of human labor, and thereby is usually not feasible for training a real-scene OPS model