Real-time semantic segmentation network for urban scenes based on global attention mechanism

In recent years, deep learning techniques have been introduced to the field of autonomous driving and have evolved with the technology to cover all aspects of autonomous driving. Researchers have begun to explore lightweighting image semantic segmentation networks and applying them to road traffic scenarios. However, existing semantic segmentation networks are usually studied based on high-resolution images, and the algorithms are difficult to be deployed on devices with limited hardware resources due to the need for more resources to compute a large number of parameters. To address this problem, this paper proposes GPIDNet, a lightweight model based on the global attention mechanism, which aims to optimize the computational efficiency while maintaining the performance. Notably, GPIDNet is able to obtain 72% mIoU and 57.9 FPS forward inference speed with a low-resolution input size of 256x256 on the Cityscapes dataset.