Trilateral dual-resolution real-time semantic segmentation network for road scenes

In the face of complex and changing real-world road scenarios, two-branch network structure can maintain good accuracy while satisfying real-time conditions. However, choice of fusing only at the end of two branches and direct fusion of low-level and high-level features leads to the waste of various information at each stage, thus limiting performance improvement of existing models. To solve this problem, a trilateral dual-resolution real-time semantic segmentation network(TDRNet) was proposed. The first branch is maintained at high resolution, while the second branch is downsampled several times, ensuring full fusion of spatial and semantic information at different stages by compact bilateral connections between the two branches. The third branch selectively learns semantic information from the second branch and fuses the other two branches in a weight-guided manner. TDRNet achieved 75.5%mIoU at 139.7 FPS on CamVid test set and 76.1% mIoU at 155.9 FPS on Cityscapes validation set. Experiments on several road datasets demonstrate that the method achieves a better balance between accuracy and real-time performance compared to other real-time semantic segmentation algorithms.