Survey on Recent Progresses of Semantic Image Segmentation with CNNs

Convolutional neural networks (CNNs) have been the mainstream in many computer vision tasks, such as image classification, object detection, face recognition and so on. We survey the state-of-the-art results on Pascal VOC 2012 semantic segmentation challenge which has made great progresses in 2015. We investigate the effectiveness of the new layers, structures and strategies behind these results proposed to produce more refined segmentation. Their main contributions focus on utilizing more structures and contextual information in the image or feature spaces. Most of these approaches serve for several independent stages in semantic image segmentation. In this paper, we discuss possible architectures to incorporate existing structures and strategies. Finally possible directions on enhancing CNNs to segment given semantic objects are proposed.