A Physically-Motivated Deformable Model Based on Fluid Dynamics

A novel deformable model for image segmentation and shape recovery is presented. The model is inspired by fluid dynamics and is based on a flooding simulation similar to the watershed paradigm. Unlike most watershed methods, our model has a continuous formulation, being described by two partial differential equations. In this model, different fluids, added by placing density (dye) sources manually or automatically, are attracted towards the contours of the objects of interest by an image force. In contrast to the watershed method, when different fluids meet they may mix. When the topographical relief of the image is flooded, the interfaces separating homogeneous fluid regions can be traced to yield the object contours. We demonstrate the flexibility and potential of our model in two experimental settings: shape recovery using manual initializations and automated segmentation.