Domain-Incremental Semantic Segmentation for Traffic Scenes

Traffic scene segmentation is an important visual perception process to provide strong support for the decision-making of autonomous driving systems. The traffic scene is an open environment that is constantly changing, and the segmentation model needs to have the ability of domain incremental learning to maintain stable performance in the changing environment. The main challenges include the diversity of the traffic scene and the accumulation of forgetting about the previous traffic scenes. In this work, an adapter-based network model is proposed to solve the domain-incremental traffic scene segmentation task. Specifically, the domain-aware adapter (DAA) module is proposed, which divides the model parameters into domain-shared and domain-specific, so that the model can handle the information of all learned domains by dynamically expanding parameters. To further alleviate catastrophic forgetting, the inter-class correlation enhancement (ICE) module is proposed, which utilizes inter-class correlation to improve segmentation accuracy for single domain and knowledge transfer between domains. Extensive experimental results show that the proposed method can achieve promising results for retaining the competitive performance for both new and old domains.