DFPNet:Dislocation Double Feature Pyramid Real-time Semantic Segmentation Network

Abstract

The feature pyramid structure has almost become a standard configuration in the object detection network. This paper introduces a dislocation double feature pyramid structure and configure it in a lightweight segmentation network. We also use the classic module (atrous spatial pyramid pooling) in the segmentation network to extract rich contextual information. Our network is called DFPNet. In order to fully verify the gain of the dislocation double feature pyramid structure for network performance, we perform a wealth of experiments on different datasets (CitySpaces and CamVid) to show that DFPNet can obtain competitive results using our novel feature pyramid module. In particular, DFPNet achieves 73.1% Mean IoU(mIoU) on the CamVid validation set with only 5.5M parameters and runtime of 117 milliseconds per image on a single RTX 2080Ti. Our code and model have been open sourced at https://github.com/Fang789/pytorch_seg.