Deformable contours or snakes are energy-minimizing models for which the minima represent solutions to contour segmentation problems. They can overcome typical problems of traditional bottom-up edge segmentation (e.g. edge gaps and spurious edges) by the use of an energy function that incorporates prior knowledge about shape in addition to terms determined by image features. Once placed in image space, the contour deforms to find the most salient contour in its neighborhood, under the influence of the generated potential field. Typical uses of deformable contours are in tracking applications or as a “power assist” tool in an interactive image editing setting.
A useful way to describe deformable contours is in terms of the following features
- contour representation
- energy formulation
- contour propagation mechanism
A system for deformable contour tracking was developed for applications such as lip reading. The system is based on the Hermite contour representation which, unlike other higher-order representations, can represent both smooth and sharp contours. The system formulates a Maximum A Posteriori (MAP) criterion for the energy function and uses a dynamical programming technique to find the optimal solution for the resulting minimization problem.
Here is a video clip of the system in action for the problem of tracking facial features, i.e. lips.