An end-to-end pipeline that turns drone imagery into 3D gaussian-splat scenes segmented per class. Structure-from-motion solves the poses, gaussian-splat training fits the scene, and a text-promptable segmentation model emits instance masks per captured view. A multi-view lifting step fuses those masks onto the underlying Gaussians as per-class layers. The shipped viewer exposes the result as a fixed concept list: click "solar panel" or "ventilation unit" and the matching splats light up.
Live at benjaminhenriksson.com/splats. Source not yet public.