DDBNet: distortion-aware dual-branch network for panoramic semantic segmentation

Panoramic Semantic Segmentation (PASS) aims to achieve comprehensive 360° scene understanding from panoramic images. However, due to the geometric distortions introduced by equirectangular projection (ERP) and the complex semantic relationships in wide field-of-view scenes, effectively modeling both global contextual information and fine-grained local details remains challenging. Existing PASS methods are mostly based on single-branch architectures, which often struggle to balance global perception capability and local semantic representation. To address these issues, we propose a distortion-aware dual-branch network for panoramic semantic segmentation. Specifically, the proposed architecture decouples global contextual modeling and local semantic representation through a collaborative spatial-semantic dual-branch design. The spatial branch enhances global context perception with large receptive fields, while the semantic branch extracts fine-grained semantic features using convolutional networks. Furthermore, a distortion-aware fusion module is designed to alleviate geometric distortions and promote effective cross-branch feature interaction. Building upon the proposed architecture, we further construct a teacher–student knowledge distillation framework. The lightweight student model is designed for efficient inference, while the teacher model leverages the complementary advantages of CNN and Transformer to provide stronger semantic supervision, thereby improving the segmentation performance of the student model without introducing additional inference overhead. Extensive experiments on multiple panoramic semantic segmentation benchmarks demonstrate the effectiveness of our approach: the teacher model achieves strong performance, while the lightweight student model maintains competitive accuracy with high inference efficiency.