What is the role of shape modeling/prior during the training

Sorry, it is difficult for me to follow the methodology. As the Bayesian model is learned from the images without occluded objects, what is the role of shape modeling/prior during training? For initializing the parameters in the Bayesian model? Or is the shape prior used for each training iteration?