Comparing global and interest point descriptors for similarity retrieval in remote sensed imagery

We investigate the application of a new category of low-level image descriptors termed interest points to remote sensed image analysis. In particular, we compare how scale and rotation invariant descriptors extracted from salient image locations perform compared to proven global texture features for similarity retrieval. Qualitative results using a geographic image retrieval application and quantitative results using an extensive ground truth dataset show that interest point descriptors support effective similarity retrieval in large collections of remote sensed imagery.