Title: Enhancing Visual Grounding in Remote Sensing through Cluster-Guided Refinement and Model Ensemble Voting

Abstract: Visual grounding, a critical element of interpretable vision systems, involves identifying image areas that match natural language descriptions. However, applying this technique to remote sensing imagery presents unique difficulties, including intricate scenes, diminutive objects, and significant scale fluctuations. Because a solitary model often struggles to handle such varied complexities, we introduce two distinct grounding frameworks: Sequential Grounding Refinement (SGR) and Cluster-Aware Grounding Refinement (CGR). These methods leverage the synergistic advantages of RemoteSAM, a grounding model tailored for remote sensing, and SAM3, a robust general-purpose segmentation tool. Initially, RemoteSAM generates a preliminary estimate of object positioning, which is subsequently optimized by SAM3 to yield segmentations that are both more precise and spatially coherent. Furthermore, we investigate an ensemble technique employing majority voting across six varied grounding pipelines, each possessing unique strengths. This multi-model architecture bolsters system robustness and markedly boosts localization precision. Our experiments confirm that the suggested pipelines and ensemble strategy surpass the performance of individual models, resulting in more dependable and accurate visual grounding outcomes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC