RTSNet: Real-Time Semantic Segmentation Network For Outdoor Scenes

Semantic segmentation technique plays an important role in robotics related applications, especially autonomous driving and assisted driving. Real-time semantic segmentation has very significant practical meaning, but many studies focus on accuracy, not computationally efficient solutions. In this paper, a real-time semantic segmentation network based on encoder-decoder architecture is proposed. This framework's encoder part adopted a lightweight network architecture for feature extraction and this architecture is mainly based on the MobilenetV2. Its decoder part is decided to use the Skip architecture. This architecture can utilize higher resolution feature mapping to provide adequate accuracy and greatly improve computational efficiency. We evaluated RTSNet on the Cityscapes dataset for urban scenes and compared with the state of the art real-time semantic segmentation networks. The mean intersection -over-union it can achieve on the Cityscapes dataset is about 62.0%, while it achieved 14.0 fps on NVIDIA Jetson TX2 with 360x640 input images.