Object instance recognition approaches based on the bag-of-words model are severely affected by the loss of spatial consistency during retrieval. As a result, costly RANSAC verification is needed to ensure geometric consistency between the query and the retrieved images. A common alternative is to inject geometric informa- tion directly into the retrieval procedure, by endowing the visual words with additional information. Most of the existing approaches in this category can efficiently handle only restricted classes of geometric transfor- mations, including scale and translation. In this pa- per, we propose a simple and efficient scheme that can cover the more complex class of full affine transforma- tions. We demonstrate the usefulness of our approach in the case of planar object instance recognition, such as recognition of books, logos, traffic signs, etc.