RelationNet: Learning Deep-Aligned Representation for Semantic Image Segmentation

Abstract

Semantic image segmentation, which assigns labels in pixel level, plays a central role in image understanding. Recent approaches have attempted to harness the capabilities of deep learning. However, one central problem of these methods is that deep convolutional neural network gives little consideration to the correlation among pixels. To handle this issue, in this paper, we propose a novel deep neural network named RelationNet, which utilizes CNN and RNN to aggregate context information. Besides, a spatial correlation loss is applied to train RelationNet to align features of spatial pixels belonging to same category. Importantly, since it is expensive to obtain pixel-wise annotations, we exploit a new training method to combine the coarsely and finely labeled data. Experiments show the detailed improvements of each proposal. Experimental results demonstrate the effectiveness of our proposed method to the problem of semantic image segmentation, which obtains state-of-the-art performance on the Cityscapes benchmark and Pascal Context dataset.