From images to maps

This paper proposes a two-stage approach to reconstruct the map of a scene given tagged photographs of that scene. In the first stage, several methods are proposed that transform tag data from photographs into an intermediate distance matrix. These methods are compared against each other. In the second stage, a method based on the physical mass-spring system is proposed that transforms the distance matrix into a map. It is compared against and outperforms MDS-MAP(P) given human-tagged input photographs. Experiments are carried out on two test datasets. Ana evaluation method is described and the optimal overall reconstruction generates maps with accuracies of 47% and 66% respectively for the datasets, both scoring 40% higher than the average random reconstruction. The proposed approach is applied to three sample datasets, and upon a qualitative inspection, the resulting maps show that they are able to convey the organization of the scenes at a high level.