CRF with locality-consistent dictionary learning for semantic segmentation

The use of top-down categorization information in bottom-up semantic segmentation can significantly improve its performance. The basic Conditional Random Field (CR-F) model can capture the local contexture information, while the locality-consistent sparse representation can obtain the category-level priors and the relationship infeature space. In this paper, we propose a novel semantic segmentation method based on an innovative CRF with locality-consistent dictionary learning. The framework aims to model the local structure in both location and feature space as well as encourage the discrimination of dictionary. Moreover, an adapted algorithm for the proposed model is described. Extensive experimental results on Graz-02, PASCAL VOC 2010 and MSRC-21 databases demonstrate that our method is comparable to or outperforms state-of-the-art Bag-of-Features (BoF) based segmentation methods.