G3AR: Graph-Guided Neural Visual Geometry for Scalable Multi-Sequence Aerial Registration

G3AR scales feed-forward neural visual geometry to large, irregular aerial collections. A geometrically verified proximity graph organizes overlapping image chunks, which are processed independently and aligned through shared-image Sim(3) transformations. Across four real scenes, the framework improves pose accuracy and runtime in matched VGGT- and Pi3-backed comparisons, while its DA3-backed variant achieves the lowest pose error among the evaluated neural-geometry methods.