A continuous formulation of Conditional Random Fields (CRFs)



Dana Cobzas,University of Alberta
Mark Schmidt,UBC


References

Cobzas D. and Schmidt M. Increased Discrimination in Level Set Methods with Embedded Conditional Random Fields, IEEE Computer Vision and Pattern Recognition (CVPR) 2009


Mark's CVPR poster


Description

We propose a novel approach for improving level set segmentation methods by embedding the potential functions from a discriminatively trained conditional random field (CRF) into a level set energy function. The CRF terms can be efficiently estimated and lead to both discriminative local potentials and edge regularizers that take into account interactions among the labels. Unlike discrete CRFs, the use of a continuous level set framework allows the natural use of flexible continuous regularizers such as shape priors.
Below are some promising results for the method on two difficult medical image segmentation tasks.


Results on 3D MRI brain tumour segmentation data

Results on 2D CT muscle segmentation data where the level set methods use a shape prior