SSRNet: A Two-branch Real-time Semantic Segmentation Network for Road Scenes

Abstract:
Semantic segmentation, one of the fundamental tasks in computer vision, requires classification for each pixel in an image, and thus semantic segmentation is time-consuming. With the rise of technologies such as autonomous driving, the realtime capability of semantic segmentation is becoming more and more important. In this paper, the authors propose a novel two-way real-time semantic segmentation network (SSRNet). We create an efficient feature extraction residual block using channel split and channel shuffle to balance operational efficiency with feature extraction and design a shared spatial path. To improve the segmentation performance of the model, we designed a feature fusion module with a skip connection. Evaluated on a single 1080Ti GPU, SSRNet can run at over 85 FPS and achieve 71.2% MIoU on the Cityscapes dataset.