Fast Structured Prediction Using Large Margin Sigmoid Belief Networks

Abstract

Images usually contain multiple objects that are semantically related to one another. Mapping from low-level visual features to mutually dependent high-level semantics can be formulated as a structured prediction problem. Current statistical models for structured prediction make simplifying assumptions about the underlying output graph structure, such as assuming a low-order Markov chain, because exact inference becomes intractable as the tree-width of the underlying graph increases. Approximate inference algorithms, on the other hand, force one to trade off representational power with computational efficiency. In this paper, we present large margin sigmoid belief networks (LMSBNs) for structured prediction in images. LMSBNs allow a very fast inference algorithm for arbitrary graph structures that runs in polynomial time with high probability. This probability is data-distribution dependent and is maximized in learning. The new approach overcomes the representation-efficiency trade-off in previous models and allows fast structured prediction with complicated graph structures. We present results from applying a fully connected model to semantic image annotation, image retrieval and optical character recognition (OCR) problems, and demonstrate that the proposed approach can yield significant performance gains over current state-of-the-art methods.