Identifying related landmark tags in urban scenes using spatial and semantic clustering

There is considerable interest in developing landmark saliency models as a basis for describing urban landscapes, and in constructing wayfinding instructions, for text and spoken dialogue based systems. The challenge lies in knowing the truthfulness of such models; is what the model considers salient the same as what is perceived by the user? This paper presents a web based experiment in which users were asked to tag and label the most salient features from urban images for the purposes of navigation and exploration. In order to rank landmark popularity in each scene it was necessary to determine which tags related to the same object (e.g. tags relating to a particular café). Existing clustering techniques did not perform well for this task, and it was therefore necessary to develop a new spatial-semantic clustering method which considered the proximity of nearby tags and the similarity of their label content. The annotation similarity was initially calculated using trigrams in conjunction with a synonym list, generating a set of networks formed from the links between related tags. These networks were used to build related word lists encapsulating conceptual connections (e.g. church tower related to clock) so that during a secondary pass of the data related network segments could be merged. This approach gives interesting insight into the partonomic relationships between the constituent parts of landmarks and the range and frequency of terms used to describe them. The knowledge gained from this will be used to help calibrate a landmark saliency model, and to gain a deeper understanding of the terms typically associated with different types of landmarks.