Generalizing Geometry-Guided Mamba as a Plug-and-Play Context Module for CNN-based Semantic Segmentation
Researchers have adapted a geometry-guided Mamba model, originally from DGM-Net, to serve as a plug-and-play context module for CNN-based semantic segmentation. This approach injects geometric guidance into the selective scan process, enabling long-range feature propagation modulated by boundary and centripetal-flow cues. When integrated into six different CNN segmentation models, the geometry-guided SSM modules consistently improved mean Intersection over Union (mIoU) scores on the Cityscapes dataset with only a slight increase in computational cost. AI
IMPACT Enhances existing CNN segmentation models with improved context aggregation, potentially leading to more accurate image analysis in computer vision tasks.