With the development of autonomous driving and general real-time vision applications, semantic segmentation is increasingly expected to provide reliable pixel-level scene understanding under strict latency and computational constraints. Existing real-time semantic segmentation methods still struggle to effectively balance contextual semantics and local details during feature fusion. This difficulty may lead to boundary blur and the loss of fine-grained structures, and is mainly related to insufficient selectivity and boundary friendliness in the fusion process. To address this issue, the Edge-Guided Residual Spatial Gating Fusion Network (ERSGNet)is proposed as a three-branch decoupled framework for real-time semantic segmentation. In the ERSGNet, detail features Fd, contextual features Fc, and edge features Fe are separately extracted to reduce mutual interference among different types of information. Spatially Gated Interaction (SGI) is introduced to inject contextual semantics into the detail branch in a pixel-wise selective manner, the Parallel Context Pyramid (PCP) is used to enhance multi-scale contextual representation, and Edge-Guided Residual Spatial Gating Fusion (ERSG-Fuse) performs edge-guided adaptive fusion between Fd and Fc under the guidance of Fe. The ERSGNet is evaluated on the CamVid and Cityscapes datasets. On CamVid, the ERSGNet achieves 73.33% mean Intersection-over-Union (mIoU) at 166.27 frames per second (FPS). On Cityscapes, the ERSGNet achieves 69.21% mIoU at 53.93 FPS. These results suggest that the ERSGNet achieves a favorable accuracy–efficiency trade-off while supporting boundary-related and fine-grained structural representation.
Wu et al. (Tue,) studied this question.