Image Modeling using Tree Structured Conditional Random Fields

In this paper we present a discriminative framework based on conditional random fields for stochastic modeling of images in a hierarchical fashion. The main advantage of the proposed framework is its ability to incorporate a rich set of interactions among the image sites. We achieve this by inducing a hierarchy of hidden variables over the given label field. The proposed tree like structure of our model eliminates the need for a huge parameter space and at the same time permits the use of exact and efficient inference procedures based on belief propagation. We demonstrate the generality of our approach by applying it to two important computer vision tasks, namely image labeling and object detection. The model parameters are trained using the contrastive divergence algorithm. We report the performance on real world images and compare it with the existing approaches.