GLoBa: Global-Local Fusion MamBa for RGB-D Semantic Segmentation

RGB-D semantic segmentation benefits from combining complementary RGB and depth (D) images but is challenged by their differing characteristics. Mamba offers a promising fusion paradigm, yet its sequential processing can disrupt 2D spatial structure and overlook local dependencies. To address this limitation and better facilitate multi-modal feature fusion, we propose Global-Local Fusion Mamba (GLoBa), which uses a dual-branch encoder for multi-scale feature extraction, a global-local fusion module with multi-scale local scanning strategy, and a decoder to capture both global and local information. Experiments on SUN RGB-D and NYU Depth V2 demonstrate the effectiveness of our approach.