Parallel Complement Network for Real-Time Semantic Segmentation of Road Scenes

Real-time semantic segmentation is in intense demand for the application of autonomous driving. Most models tend to use large feature maps and complex structures enhance representation power high accuracy. However, these inefficient designs increase amount computational costs, which hinders model be applied on In this paper, we propose a lightweight real-time model, named Parallel Complement Network (PCNet), address challenging task with fewer parameters. A layer introduced generate complementary features receptive field. It provides ability overcome problem similar encoding among different classes, further produces discriminative representations. With inverted residual structure, design block construct proposed PCNet. Extensive experiments are carried out road scene datasets, i.e., CityScapes CamVid, make comparison against several state-of-the-art models. The results show that our has promising performance. Specifically, PCNet* achieves 72.9% Mean IoU using only 1.5M parameters reaches 79.1 FPS $1024\times 2048$ resolution images GTX 2080Ti. Moreover, system best accuracy when being trained from scratch.