Very Fast Semantic Image Segmentation Using Hierarchical Dilation and Feature Refining

With the rapid development of deep learning techniques, semantic image segmentation has been considerably improved recently, which is viewed as the key problem of scene understanding in computer vision. These advances are built upon the capability of complex architectures for deep neural network. In this paper, we present a novel deep neural network architecture designed for semantic image segmentation. In order to improve the segmentation accuracy, we introduce a novel hierarchical dilation block to effectively enlarge the size of receptive field and enable multi-scale processing in fully convolutional neural network. Moreover, we exploit the technique of bypass and intermediate supervision to capture the context information during upsampling and refining coarse features. We have conducted extensive experiments on several popular semantic segmentation testbeds, including Cityscapes, CamVid, Kitti, and Helen facial datasets. The experimental results demonstrate that our proposed approach runs two times faster than the state-of-the-art method. Our full system is able to obtain realtime inference performance on 1080P images using a PC with single GPU. It executes a network forwarding at 200fps in our experiment while retaining high accuracy. Our proposed approach not only runs faster than the existing realtime methods but also performs on par with them.