Skip to Main Content
One commonly used approach to scene localization and landmark recognition is to match an input image against a large annotated database of images using local image features. However problems exist with these approaches relating to memory constraints and the processing time required to compare high dimensional image feature vectors in a very large scale database. We investigate a new landmark classification technique which takes advantage of the fact that there is considerable overlap in visually similar images of landmarks in any large public photo repository. A large number of images containing landmarks are clustered into visually similar clusters. Classification models are then implemented and trained based on global histograms of interest point features from these clusters to create models which can be used for robust real-time accurate classification of images containing these landmarks. We also investigate different techniques for the creation of these classification models to ascertain how best to guarantee a high level of robustness, accuracy and speed.