A pictorial structure modelfor an object is given by a collection of parts with connections between certain pairs of parts. The framework is quite general, in the sense that it is independent of the specific scheme used to model the appearance of each part as well as the type of connections between parts. A natural way to express such a model is in terms of an undirected graph G = (V, E), where the vertices V = {v1 , . . . , vn} correspond to the n parts, and there is an edge (vi , vj) ∈ E for each pair of connected parts vi and vj . An instance of the object is given by a configuration L = (l1 , . . . , ln), where each li specifies the location of part vi. The location of each part can simply specify its position in the image, but more complex parameterizations are also possible. For example, for the person model the location of a part specifies a position, orientation and scale. In the problem of matching a pictorial structure to an image is defined in terms of an energy function to be minimized. The cost or energy of a particular configuration depends both on how well each part matches the image data at its location, and how well the relative locations of the parts agree with the deformable model.
Chat with our AI personalities