Interactive Image Search with Attribute-based Guidance and Personalization

Interactive image search, where a user initiates a query with keywords and refines it via feedback, can be enhanced using attributes. To minimize the user’s effort, we propose to let the system guide a user’s attribute-based compara-tive feedback with “pivot ” exemplars that most reduce the system’s uncertainty. Since humans vary in how they per-ceive the link between a named property and image content and might disagree on an attribute’s presence, we further show how to efficiently learn user-specific attribute models. We demonstrate that attribute adaptation and system-driven feedback allow the user to quickly find his desired target. 1