It is often useful to represent a single example by a set of the local features that comprise it. However, this representation poses a challenge to many conventional learning techniques, since sets may vary in cardinality and the elements are unordered. To compare sets of features, researchers often resort to solving for the least-cost correspondences, but this is computationally expensive and becomes impractical for large set sizes. We have developed a general approximate matching technique called the pyramid match that measures partial match similarity in time linear in the number of feature vectors per set. The matching forms a Mercer kernel, making it valid for use in many existing kernel-based learning methods. We have demonstrated the approach for various learning tasks in vision and text processing, and find that it is accurate and significantly more efficient than previous approaches.
Subjects: 12. Machine Learning and Discovery; 19. Vision
Submitted: Apr 24, 2007