Analysing the Structure of Semantic Concepts in Visual Databases

In this paper we study how the Self-Organizing Map (SOM) can be used in analysing the structure of semantic concepts in visual data. We investigate two data sets with concept labels provided by humans, and unlabelled data for which we utilise automatically detected concept membership scores by using models trained on a labelled data set. By arranging the concept memberships of visual objects as components of a vector, they can be used as the feature space for training a SOM. A visual and qualitative analysis of the SOM distributions of different concepts is augmented with a quantitative analysis based on measuring the Earth Mover’s Distance between the vector distributions on the 2D SOM surface. In particular we study the PASCAL VOC 2007 and TRECVID 2010 databases, which are two large image and video data sets.