Splitting crowded views helps remote sensing models count
GeoScope: Training-Free Program Execution beyond the Counting Horizon for Ultra-High-Resolution Remote Sensing Question Answering · arXiv:2611.12699
In plain words
Large aerial images show many tiny objects. Even after a model looks at the right part, it can miss objects when that view is crowded.
GeoScope turns a question into steps for choosing image regions. It asks a vision-language model to count locally, splits views whose counts exceed the model’s measured limit, and calculates directions from object locations.
On held-out tests of regional counting and object relationships, it improved Qwen3-VL-8B’s accuracy.
The research
For a count from a large image, choosing the right patch is not enough: a model may still miss objects inside it.
GeoScope changes the expectation from “zoom in and trust the count” to checking whether a view is crowded and splitting it when needed; it also calculates directions from object locations.
GeoScope turns questions into image operations, splitting crowded counting views and calculating directions from object locations without updating either model.
On XLRS-Bench, which tests counts in marked image regions and directions between objects, accuracy measures the share of questions answered correctly. With the same Qwen3-VL-8B backbone and held-out questions, GeoScope improved accuracy on both tasks over direct whole-image answers.
The paired results support using question-guided views and explicit spatial calculations for these tasks, while the counting tests on another image collection also favored GeoScope.
The counting limit and split-response model were measured on aerial and drone images with a few object classes. Their behavior on crowds, cells, or everyday scenes remains unresolved.
What the evidence shows
- A method built around image operations
GeoScope is a training-free method that turns a question about an object or relationship into image operations. A text-only language model writes the operations from the question, while a deterministic executor uses a frozen vision-language model for local observations.
The executor subdivides crowded counting views and computes directions from object locations. This assigns counting and spatial calculations different roles instead of relying on one image-wide answer.
- A view can still be too crowded
The key diagnostic is that selecting a view does not make every count reliable. On DOTA aerial-image crops, most of eight tested open models began systematically undercounting beyond about five objects in a view, though the onset varied by model.
For Qwen3-VL-8B, mean signed relative error reached −57% above 160 objects. The error also worsened in larger views where objects appeared smaller, so the measured limit depends on the view and imagery.
- The measured limit guides splitting
The paired comparison suggests that GeoScope’s question-guided views and explicit spatial calculations help Qwen3-VL-8B answer both regional-counting and object-relationship questions. This evidence is limited to those XLRS-Bench tasks and that backbone.
- Counting gains transfer to another test
On UHR-Micro, GeoScope also improved Qwen3-VL-8B’s counting scores against a zero-shot rerun, with a larger gain on regional than global counting.
Because the score gives partial credit for counts closer to the true number, this indicates better average counting under that measure; a comparison using recorded unsplit-view counts attributes much of the global-counting gain to subdivision in that setup.
- One test has a misleading option
MME-RealWorld counting comparisons are difficult to interpret because a rarely correct answer option is selected often by the tested backbones. The authors report that this option bias distorts comparisons between counting methods on this test.
Deep
- Original paper authors
- Guohui Deng, Yuhang Jiang, Chengjun Wang, Zhongwen Hu, Zhaocong Wu
- Submitted to arXiv
- License
- arXiv non-exclusive