Interactive Visualization of Multidimensional Feature Spaces

Image similarity models characterize images as points in high-dimensional feature spaces. Each point is represented by a combination of distinct features, such as brightness, color histograms or texture characteristics of the image, etc. For the design and tuning of features, and thus the effectiveness of the image similarity model, it is important to understand the interrelations of individual features and the implications on the structure of the feature space. In this paper, we discuss an interactive visualization tool for the exploration of multidimensional feature spaces. Our tool uses a graph as an intermediate representation of the points in the feature space. A mass spring algorithm is used to layout the graph in a 2D space in which arrangements of similar images are attracted to each other and dissimilar images are repelled. The emphasis of the visualization tool is on interaction: users may influence the layout by interactively scaling dimensions of the feature space. In this way, the user can explore how a feature behaves in relation to other features.