Remote Sensing Visual Grounding through Diffusion Model

Remote Sensing Visual Grounding (RSVG) focuses on identifying location of targets in RS images mentioned by text descriptions. Existing method typically rely on multi-layer transformer framework, which uses visual encoder, text encoder and fusion decoder to align visual and text features. However, the absence of continuous and progressive text guidance can lead to inconsistencies between visual and textual features, thereby limiting grounding ability. Thus, in this paper, we present a novel remote sensing visual grounding framework by diffusion model, RSD-VG, which use U-Net to progressively refine queried object boxes by denoising using consistent language guidance. We validate the efficiency of RSD-VG on DIOR-RSVG dataset and compare it with previous methods and achieve promising results.