A Fully-Convolutional Framework for Semantic Segmentation

Abstract

In this paper we propose a deep learning technique to improve the performance of semantic segmentation tasks. Previously proposed algorithms generally suffer from the over-dependence on a single modality as well as a lack of training data. We made three contributions to improve the performance. Firstly, we adopt two models which are complementary in our framework to enrich field-of-views and features to make segmentation more reliable. Secondly, we repurpose the datasets form other tasks to the segmentation task by training the two models in our framework on different datasets. This brings the benefits of data augmentation while saving the cost of image annotation. Thirdly, the number of parameters in our framework is minimized to reduce the complexity of the framework and to avoid over- fitting. Experimental results show that our framework significantly outperforms the current state-of-the-art methods with a smaller number of parameters and better generalization ability.