From Visual Representation to Object Discovery and Back

Abstract

This thesis summarizes the author’s post-PhD work that has a major focus on instance-level recogni- tion tasks, such as instance-level search primarily, but also instance-level recognition. Visual representa- tion is at the core of most computer vision tasks. This is addressed in the first part of the manuscript with crafting and learning of visual representations and similarity measures. A number of different represen- tation approaches are summarized under the same framework provided by a match kernel formulation, while the interplay between local and global representation is highlighted. Given the representation, efficient estimation of similarity with respect to a large number of examples enables visual search appli- cations. This is covered in the second part especially in the form of query expansion, which goes beyond pairwise similarity and treats a collection of examples as a whole. The last part discusses object discovery as an outcome of exploiting visual similarity and search within a large unordered collection of examples. Its result is then used to improve other components, namely representation and search, therefore, forming a circle and pronouncing the synergy between different contributions in this thesis.