A new robust context-based dense CRF model for image labeling

Fully-connected conditional random fields (CRF) models have recently been developed for image labeling task to incorporate interactions of all pairs of pixels in the image. Efficient inference in fully-connected models is very sensitive to initialization of the unary potentials. In this paper, we propose a new robust context-based fully-connected CRF model which alleviates initialization sensitivity of inference in dense CRFs. The new model integrates an extra hidden node that accounts for the overall context of the image and is connected to all other pixel nodes. By incorporating the new context node in CRF graph, we maximize probability of labeling configuration jointly with the image´s context. Therefore, wrong initializations of objects that contradict the overal context could be refined. We define the context-based unary and pairwise potentials and further derive the inference algorithm for the proposed model based on the mean field approximation method. We run experiments over the benchmark MSRC image database and demonstrate that the new model improves object recognition accuracy by about 21%. We show where the conventional CRF model is impeded by wrong initialization of unary potentials, the proposed model identifies the labels correctly.