A Co-occurrence Prior for Continuous Multi-label Optimization

Abstract

Abstract. To obtain high-quality segmentation results the integration of semantic information is indispensable. In contrast to existing segmentation methods which use a spatial regularizer, i.e. a local interaction between image points, the co-occurrence prior [15] imposes penalties on the co-existence of different labels in a segmentation. We propose a continuous domain formulation of this prior, using a convex relaxation multi-labeling approach. While the discrete approach [15] is employs minimization by sequential alpha expansions, our continuous convex formulation is solved by efficient primal-dual algorithms, which are highly parallelizable on the GPU. Also, our framework allows isotropic regularizers which do not exhibit grid bias. Experimental results on the MSRC benchmark confirm that the use of co-occurrence priors leads to drastic improvements in segmentation compared to the classical Potts model formulation when applied .

Optimizing the Relevance-Redundancy Tradeoff for Efficient Semantic Segmentation

“…Semantic segmentation -sometimes also referred to as class-specific segmentation -aims at jointly computing a partitioning of the image plane and a semantic labeling of the various regions in terms of previously learned object classes. Numerous works are focused on the development of sophisticated regularizers for this problem: co-occurrence priors [9,18] have been suggested to learn and penalize the joint occurrence of semantic labels within the same image. Proximity priors [1] have been proposed to penalize the co-occurrence of labels within a certain spatial neighborhood.…”