Revealing what to extract from where, for object-centric content based image retrieval (CBIR)

Content Based Image Retrieval (CBIR) techniques retrieve similar digital images from a large database. As the user often does not provide any clue (indication) of the region of interest in a query image, most methods of CBIR rely on a representation of the global content of the image. The desired content in an image is often localized (e.g. car appearing salient in a street) instead of being holistic, demanding the need for an object-centric CBIR. We propose a biologically inspired framework WOW ("What"Object is "Where") for this purpose. Design of WOW framework is motivated by the cognitive model of human visual perception and feature integration theory (FIT). The key contributions in the proposed approach are: (i) Feedback mechanism between Recognition ("What") and Localization ("Where") modules (both supervised), for a cohesive decision based on mutual consensus; (ii) Hierarchy of visual features (based on FIT) for an efficient recognition task. Integration of information from the two channels ("What" and "Where") in an iterative feedback mechanism, helps to filter erroneous contents in the outputs of individual modules. Finally, using a similarity criteria based on HOG features (spatially localized by WOW) for matching, our system effectively retrieves a set of rank-ordered samples from the gallery. Experimentation done on various real-life datasets (including PASCAL) exhibits the superior performance of the proposed method.