DSMRSeg: Dual-Stage Feature Pyramid and Multi-Range Context Aggregation for Real-Time Semantic Segmentation

Real-time semantic segmentation is a challenging task in computer vision. Many researches emphasize real-time inference speed while neglecting segmentation quality. To tackle this problem, we propose a framework called DSMRSeg to achieve high-speed with high-accuracy result after training on only one GPU. Overall, we accomplish this by three core components: (1) Dual-Stage Feature Pyramid Network structure is designed to obtain richer multi-scale information and enhance the entire features hierarchy by bidirectionally propagating features with strong semantics and accurate localization. (2) Multi-Range Context Module is developed to expand receptive fields by aggregating the local dense features and multi-range context information. (3) Light-weight Feature Fusion Module is proposed to merge dual-stage features effectively. We evaluate DSMRSeg on Cityscapes, CamVid and BDD100K datasets and produce competitive results compared with the state-of-the-art methods. Specifically, DSMRSeg achieves 75.5% mIoU on Cityscapes test set, with speed of 40 FPS on one NVIDIA GTX1080 card for 1024 × 512 high-resolution image.