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

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