Manipulating Bilevel Feature Space for Category-Aware Image Exploration

The demand for interactively designing the image feature space has been increasing due to the ongoing need for image retrieval, recognition, and labeling. Although conventional methods provide an interface for locally rearranging such a feature space, category-level global manipulation is still missing and thus manually rearranging the overall image categorization usually requires a time-consuming task. This paper presents a novel approach to exploring images in the database through the manipulation of bi-level feature space representations, where the upper-and lower-level representations characterize the global categories and local features of the images, respectively. In this approach, the upper-level space describes similarity relationship among the underlying categories extracted from the bag-of-features model, while the lower-level space encodes the closeness between a pair of images within the same category. The key idea behind this approach is to associate the relationship between the two feature spaces with a two-layered graph representation and project it onto 2D screen space using pivot MDS for user manipulation. Experimental results are provided to demonstrate that our approach allows users to understand the entire structure of the given image dataset and reorganize the layout according to their preference both locally and globally.