Higher Order Potentials in End-to-End Trainable Conditional Random Fields

摘要: We tackle the problem of semantic segmentation using deep learning techniques. Most systems include a Conditional Random Field (CRF) model to produce structured output that is consistent with visual features image. With recent advances in learning, it becoming increasingly common perform CRF inference within neural network facilitate joint pixel-wise Convolutional Neural Network (CNN) classifier. While basic CRFs use only unary and pairwise potentials, has been shown addition higher order potentials defined on cliques more than two nodes can result better outcome. In this paper, we show types potential, namely, object detection based superpixel be included embedded network. design these allow efficient differentiable mean-field algorithm, making possible implement our as stack layers As result, all parameters richer jointly learned CNN classifier during end-to-end training entire find significant improvement results introduction trainable potentials.