Author Description

"A visual instance is a visually unique entity (e.g., Brooklyn bridge), or a set of entities with identical visual appearance and hence not visually distinguishable (e.g., a Miffy doll with many identical copies). The same instance may vary tremendously in appearance in different images due to self-deformation, occlusion, illumination variations, viewpoint change and other recording factors. This thesis is dedicated to visual instance search from one example. Given one image of an instance, the goal is to retrieve all the images of the same instance, regardless of the appearance variations in different recordings. The main findings are as follows. 1) Localized search by evaluating multiple local boxes in the image is advantageous, as the spatial extent of an instance is often limited to a portion of an image and the signal-to-noise ratio within the box is higher than in the entire image. 2) A strict matching on local details is helpful for one-sided instances with limited viewpoint variations, e.g., logos, but not suited for multi-sided instances like shoes, when photographed from every possible viewing angle. 3) The attribute-based method, which learns a set of visual aspects, is generically applicable to both one-sided and multi-sided instances. The learned aspects are invariant to recording factors and discriminative among different instances. These aspects are generalizable to previously unseen instances. 4) Tracking, when provided with a powerful similarity metric, can be addressed as an instance search problem where the initial target in the starting frame is considered as the query."--Samenvatting auteur.