GeoScope: Training-Free Program Execution beyond the Counting Horizon for Ultra-High-Resolution Remote Sensing Question Answering
All eight tested vision-language models increasingly undercounted crowded aerial views, with most hitting a limit near five objects.
The preprint finds that this limit falls as views grow and objects appear smaller. GeoScope splits views only when they exceed the model’s measured counting range, uses a text-only language model to turn questions into spatial programs, and computes directions from grounded boxes. Without training, it improved Qwen3-VL-8B by 16.3 percentage points on regional counting and 10.5 points on spatial relations in XLRS-Bench. The authors report the highest accuracy on regional counting and across the benchmark by question, including comparisons with trained methods.
ArXiv · AI/CL/LG's note
The preprint finds that this limit falls as views grow and objects appear smaller. GeoScope splits views only when they exceed the model’s measured counting range, uses a text-only language model to turn questions into spatial programs, and computes directions from grounded boxes. Without training, it improved Qwen3-VL-8B by 16.3 percentage points on regional counting and 10.5 points on spatial relations in XLRS-Bench. The authors report the highest accuracy on regional counting and across the benchmark by question, including comparisons with trained methods.
ArXiv · AI/CL/LG's note
score 4