Learning Spatial Configuration Models Using Modified Dirichlet Priors

Semantic scene classification is a challenging problem in computer vision. Special-purpose semantic object and material (e.g., sky and grass) detectors help, but are faulty in practice. In this paper, we propose a generative model of outdoor scenes based on spatial configurations of objects in the scene. Because the number of semantically-meaningful regions (for classification purposes) in the image is expected to be small, we infer exact probabilities by utilizing a brute-force approach. However, it is impractical to obtain enough training data to learn the joint distribution of the configuration space. To help overcome this problem, we propose a smoothing technique that modifies the naive uniform (Dirichlet) prior by using modelbased graph-matching techniques to populate the configuration space. The proposed technique is inspired by the backoff technique from statistical language models. We compare scene classification performance using our method with two baselines: no smoothing and smoothing with a uniform prior. Initial results on a small set of natural images show the potential of the method. Detailed exploration of the behavior of the method on this set may lead to future improvements.