Emergent Properties from Feature Co-occurrence in Image Collections

This paper proposes a novel approach to explore emergent patterns in images in an unsupervised setting. We consider emergent patterns to be sets of co-occurring visual words that appear together more often than chance would indicate. Rather than focusing on finding ways to learn a large number of objects or their categories we focus on analyzing behavior associated with emergent patterns. We show that these patterns emerge from the data and in some cases relate to object identifiers. We extract SIFT descriptors [1] and then cluster them to represent each image as a bag-of-words. To encode co-occurrences between visual words we represent them as edges of a graph which are weighted by the number of images containing a particular co-occurrence. Performing a statistical analysis on weights of the edges identifies words which co-occur significantly more often than expected. These highly co-occurring nodes produce clusters in the graph which can be separated using normalized cuts. Applying normalized cuts reveals that in simple images datasets these emergent clusters can identify object classes. Results on more complex datasets like Caltech101 [2] show that interesting patterns other than object classes can also emerge from the data.