Existing local feature detectors such as Scale Invariant Feature Transform (SIFT) usually produce a large number of features per image. This is a major disadvantage in terms of the speed of search and recognition in a run-time application. Besides, not all detected features are equally important in the search. It is therefore essential to select informative descriptors. In this paper, we propose a new approach to selecting a subset of local feature descriptors. Uniqueness is used as a filtering criterion in selecting informative features. We formalize the notion of uniqueness and show how it can be used for selection purposes. To evaluate our approach, we carried out experiments in urban building recognition domains with different datasets. The results show a significant improvement not only in recognition speed, as a result of using fewer features, but also in the performance of the system with selected features.