Image semantic segmentation optimization by Conditional Random Field integrated with object clique potential

Image semantic segmentation is a pixel-wise label assigning problem. Recent image semantic segmentation methods based on Fully Convolutional Networks (FCNs) combined with Conditional Random Fields (CRFs). These systems all employ CRFs with only unary and pairwise potentials, which can merely refine simple-structured delineation situation. In this paper, we propose a more delicate CRF-inference structure to deal with delineation optimization problems, which introduces a new type of potential defined on object cliques. And by applying this method to Pascal VOC 2012 dataset, our system outperform the baseline system by 2% on mean Intersection over Union (IoU) standard.