A research team has built a way to pinpoint objects in aerial images from plain-language descriptions — no model fine-tuning required.
The system, called GeoSelect, converts a natural-language phrase like "the largest ship" or "the second court from the left" into a small spatial program, then executes it deterministically against the image. A frozen language model writes the program; a rules checker vets it; an executor runs it. The key design choice is a single "scored candidate set" abstraction that lets geometric position maps and discrete ordering operators compose cleanly — so spatial, comparative, and ordinal instructions all flow through the same pipeline. If a program fails at any step, the system falls back to a geometry-only mode rather than returning nothing.
The hard part in aerial imagery has always been spatial relationships. Most no-training methods lean on vision-language similarity scores, which can match "red roof" well enough but struggle with "third building north of the intersection." GeoSelect's explicit execution layer handles those constructions directly, and because every intermediate step is inspectable, failures are diagnosable rather than opaque.
On the RRSIS-D benchmark, GeoSelect scores 58.86 mIoU — more than double the previous best training-free method — and 55.27 mIoU on RISBench, running on a single GPU with no task-specific supervision. The authors ran ablations confirming the gain comes from the execution design, not the choice of backbone model. Code will be released when the paper clears peer review.