Semantic Segmentation Using Fully Convolutional Networks and Random Walk with Prediction Prior

Abstract

Fully Convolutional Networks (FCNs) for semantic segmentation always lead to coarse predictions, especially in border regions. Improved models of FCNs with conditional random fields (CRFs), however, cause significant increase in model complexity and scattered distribution of pixels in border regions. To address these issues, we propose a novel approach combining random walk with FCNs to capture global features and refine border regions of segmentation results. We design a double-erosion mechanism on the prediction results of FCNs to initialize random walk, and apply prediction scores as a global prior of random walk model by adding an extra item into the weight matrix of the graph constructed from an image. Experimental results show that the proposed method acts better than Dense CRF in pixel accuracy and mean IoU, and obtains smoother results. In addition, our method significantly reduces the time cost of refinement process.