This paper presents a new approach for the object categorization problem. Our model is based on the successful `bag of words' approach. However, unlike the original model, image features (keypoints) are not seen as independent and orderless. Instead, our model attempts to discover intermediate representations for each object class. This approach works by partitioning the image into smaller regions then computing the spatial relationships between all of the informative image keypoints in the region. The results show that the inclusion of spatial relationships leads to a measurable increase in performance for two of the most challenging datasets.