Semantic segmentation with modified deep residual networks

Abstract

A novel semantic segmentation method is proposed, which consists of the following three parts: (I) First, a simple yet effective data augmentation method is introduced without any extra GPU memory cost during training. (II) Second, a deeper residual network is constructed through three effective techniques: dilated convolution, LSTM network and multi-scale prediction. (III) Third, an online hard pixels mining is adopted to improve the segmentation performance. We combine these three parts to train an end-to-end network and achieve a new state-of-the- art segmentation accuracy of 79.3% on PASCAL VOC 2012 test set at the time of submission.