Space and semantics sensitive segmentation using derivative

Semantic segmentation struggles with regions exhibiting spatial and semantic incoherence, such as small objects, edges and inconsistent intra-object areas. To address these challenges, we introduce a Derivative-Guided Network (DGN) that leverages spatial and semantic derivatives to improve segmentation accuracy in such regions. Our approach features two novel modules: the Derivative-Guided Localization (DGL) module, which identifies regions with significant spatial and semantic variations, and the Derivative-Guided Enhancement (DGE) module, which dynamically enhances local and global features within these regions. By integrating DGL and DGE, our method improves segmentation performance in incoherent areas. Experiments on the CamVid datasets validate the performance achieved by our method.