Residual Inception Skip Network for Binary Segmentation

Abstract

This paper summarizes our approach to the Deep Globe Road Extraction challenge 2018. In this challenge we are tasked to find road networks from satellite images. First, we explain our U-Net type baseline model for the challenge. Second, we explain a new architecture that takes in the lessons from some of the popular approaches that we call Residual Inception Skip Net. Finally, we outline our cyclic learning rate based ensembling approach which improved the overall single model performance and the final solution for submission. Our final model increases the IoU by 3 points over the baseline.