Deformable shape models for 2D object segmentation

Given a set of images showing individual 2D instances of an object class, the goal is to learn object class deformation in 2D for segmentation automatically. Class deformation is modelled by linear combinations of basis shapes. Usually, given segmentation data and correspondences, such basis shapes can be easily learned with Principal Component Analysis. Here, we are dealing with unsegmented RGB images. We show how to learn segmentations and deformation sequentially in an iterative framework. Variations of the basic algorithm are explained, tested and compared. In order to introduce smoothness priors and data dependent pairwise terms, Graph-cut can be incorporated. The final results show that explicitly restricting segmentations by a linear subspace of shape deformation, leads to significant improvements.