Wide Residual Networks for Semantic Segmentation

In the task of object recognition, convolutional neural networks (CNNs) have achieved high performance. In addition, these CNNs are also applied to the field of semantic image segmentation. However, applying the classification models to semantic segmentation tasks has a problem, lack of global context and reduction in resolution. In this work, we propose global context module and high resolution path in order to solve above problems. By simply combining them with an existing classification model (wide residual networks), our methods yield high-accuracy segmentation models. Our proposed approaches produce competitive results, the mean intersection over union (IoU) 67.6% and global accuracy 91.1%, on CamVid test set.