Training image classifiers with similarity metrics, linear programming, and minimal supervision

Image classification is a classical computer vision problem with applications to semantic image annotation, querying, and indexing. Recent and effective generative techniques assume Gaussianity, rely on distance metrics, and estimate distributions, but are unfortunately not convex nor keep computational architecture in mind. We propose image content classification through convex linear programming using similarity metrics rather than commonly-used Mahalanobis distances. The algorithm is solved through a hybrid iterative method that takes advantage of optimization space properties. Our optimization problem uses dot products in the feature space exclusively, and therefore can be extended to non-linear kernel functions in the transductive setting.