Visualizing Object Clouds Through Energy Minimization

Abstract

When visualizing an object cloud, the pairwise similarity between an object and a central object of interest is used to determine the position of each object within the cloud. This however does not capture the semantic relationship of all the objects and it reduces the expectation of finding an object when performing visual search. To generate a semantic object cloud, we define and subsequently minimize an energy function that captures the pairwise similarity amongst all the objects within the cloud. The energy is minimized using several statistical machine learning techniques and we show that the generated layouts from such techniques outperform those of other object cloud algorithms on a variety of metrics for evaluating word and object cloud layouts.