Dual Attention Dual-Resolution Networks for Real-Time Semantic Segmentation of Street Scenes

Semantic segmentation is a crucial technology for autonomous vehicles to acquire information about their surrounding environment. To ensure that semantic segmentation has practical application value in autonomous driving and robotics, it must achieve corresponding real-time inference speeds. However, existing models either improve accuracy at the cost of high computational expense and long inference times or enhance inference speed by sacrificing resolution and multi-level detailed information, resulting in a significant drop in accuracy. In this paper, we propose a new architecture based on a bilateral segmentation network, called DADNet. We have designed a new attention mechanism to optimize feature maps and a feature fusion module with an attention mechanism to efficiently merge different features, effectively expanding the receptive field. Our method demonstrates an excellent balance between segmentation accuracy and speed on the Cityscapes and CamVid datasets. Specifically, DADNet achieves a mIoU of 78.2% at 90.5 FPS on the Cityscapes validation set using a single 2080Ti GPU. On the CamVid test set, it achieves a mIoU of 75.8% at 136.7 FPS. Our approach outperforms most state-of-the-art models while requiring less computational power.