A New Feature Pyramid Network For Road Scene Segmentation

Abstract

Road scene segmentation is of great significance in intelligent transportation system for different applications such as autonomous driving and semantic map building. Despite great progress in this field with the deep learning methods, there are still many difficulties such as robust segmentation of small objects and same type of objects with different sizes in different scenes. In this paper, we propose a new pyramid architecture for scene segmentation, which is a top-down architecture with lateral connections for multi-scale semantic feature maps building, and sufficiently incorporate the momentous global scenery prior. Besides, we also propose a novel training method, which combines the re-sampling, pixel-wise cost learning and transfer learning together, to deal with the imbalance problem. Experimental results on KITTI and Cityscapes dataset demonstrate effectiveness of the proposed method.