Scene-Adaptive 3-D Semantic Segmentation Method Based on Multiphase Edge Enhancement for Intelligent Vehicles

The 3-D semantic segmentation is crucial for enabling intelligent vehicles to accurately perceive and interpret their environment. Current state-of-the-art (SOTA) methods have achieved competitive results in indoor and typical urban traffic scenes. However, they commonly generate inaccurate edge detections, especially in those complex and variable scenarios. In order to improve edge awareness and robustness, we propose an efficient scene-adaptive method based on multiphase edge enhancement. First, the multitask segmentation module could simultaneously learn to predict edge detections and semantic segmentation labels using a two-branch structure. Second, considering the strong correlation between the two tasks, a cross-dimensional feature-fusion module is presented to implement the interaction of shared features for hierarchical supervision while retaining respective private features for jointly refining their predictions. Third, an edge-guided module (EGM) is induced to ensure the consistency on edges between the two tasks and realize the region-level edge awareness explicitly. Finally, joint learning is utilized to optimize the whole network by combining segmentation losses of two branches and region-energy loss of the EGM. We perform extensive evaluations on urban dataset SemanticKITTI and Nuscenes, off-road dataset RELLIS-3D, and our unstructured test set. The experimental results demonstrate that the proposed method has competitive performance in improving efficiency, accuracy, and scene-adaptivity.