Real-time Semantic Segmentation for Aggregating Long-Range Information and Region Understanding

With the improvement of hardware performance, semantic segmentation based on convolutional neural network has achieved wide concern for its' advantage of high accuracy. However, in practical application scenarios, such as autonomous driving, what we care about is not only the segmentation accuracy, but also the ability of the model to process information in real-time on edge devices. In this paper, a lightweight and efficient real-time semantic segmentation network was proposed, which is based on a two-way structure, aggregates detailed spatial information and high-level semantic information, integrates long-range dependencies and region-level understanding. The proposed algorithm achieves good performance. We conducted experiments based on an NVIDIA Geforce RTX 2080 SUPER graphics card and achieved 73.8% accuracy on Cityscapes dataset with a speed of 211 FPS, and 72.8% accuracy on Camvid dataset.