Auxiliary Edge Detection for Semantic Image Segmentation

Semantic segmentation is a challenging task which can be formulated as a pixel-wise classification problem. Most FCN-based methods of semantic segmentation apply simple bilinear up-sampling to recover the final pixel-wise prediction, which may lead to misclassification near the object edges. To solve this problem, we focus on the supplementary spatial details of semantic segmentation using edge information. We present an approach to incorporate the relevant auxiliary edge information to semantic segmentation features. By applying the explicit supervision of semantic boundary using intermediate features, the multi-tasks network learns features with strong inter-class distinctive ability. The attention-based feature fusion module fuses the high-resolution edge features with wide-receptive-field semantic features to sufficiently leverage the complementary information. Experiments on the Cityscapes dataset show the effectiveness of fusing intermediate edge information.