high performance computing on graphics processing units: hgpu.org

The introduction of prior knowledge into image analysis algorithms is a central challenge in computer vision. In this paper, we introduce the concept of proximity priors into semantic segmentation methods in order to penalize the proximity of certain object classes. Proximity priors are a generalization of purely global and purely local co-occurrence priors which have been introduced recently. The key idea is to consider pixels as adjacent if they are within a specified neighborhood of arbitrary size and shape. Respective penalties for the adjacency of various label pairs (the labels ‘sheep’ and ‘lion’ for example) can be learned statistically from a set of segmented images. We propose a variational approach which integrates morphological operators and derive an exact convex relaxation which can be minimized globally. Extensive numerical validations on an established semantic segmentation benchmark demonstrate that the proposed proximity priors compare favorably to existing approaches.