Modeling spatial patterns of shapes

We introduce a framework for modeling spatial patterns of shapes formed by multiple objects in an image. Our approach is graph-based where each node denotes an object and attributes of a node consist of that object's shape, position, orientation, and scale. Neighboring node are connected by edges, and they are allowed to interact in terms of their attributes/features. Similar to a Markov random field, but now applied to more sophisticated features space, the interactions are governed by energy functionals that can be internal or external. The internal energies, composed entirely of interactions between nodes, may include similarity between shapes and pose. The external energies, composed of outside influences, may include the data-likelihood term and the a-priori information about the shapes and the locations of the objects.