Supervised segmentation by pairwise interactions: do Gibbs models learn what we expect?

Gibbs random field image models with multiple translation invariant pairwise pixel interactions show promise for segmenting piecewise-homogeneous image textures because they allow learning of both the interaction structure and strengths from a given training sample. We discuss whether the learnt parameters fit our expectations with respect to discriminating the given textures. Experiments with natural textures show that the learning tends to adapt the model more to peculiarities of the training sample than to general discriminating features of the textures. Low segmentation errors for just the training image or the image containing big texture patches used for learning may mislead in predicting the errors for the test images. Texture inhomogeneities or different region statistics in the training and test images are outside the scope of the models. Thus, the textures have to meet specific constraints for using such a supervised segmentation in practice