A weight-adaptive dynamic model for shape segmentation

Physically based dynamic models are able to describe variable shapes without prior training. Their behaviour to find an object is intuitive, which facilitates corrections of false results. Expressing shape variation as physical feature, however, may be difficult because the physics of the model has little to do with the shape variation of instances of a class of objects. We present a dynamic model, which automatically adapts model parameters based on results of previous segmentations. The model was applied to artificial data and to images of leaves. Results show that the adapted model finds the correct shape more accurate than a model with preset parameters. Investigation of the parameterisation from adaptation also showed that they may be interpreted in terms of the semantics of the shape class represented.