Dataset viewer and quality inspector for the Test-3.2K evaluation set

UAV-MapVQA

Evaluating Foundation Models on Instance-to-Entity Grounding Question Answering across UAV Imagery and Vector Maps

Can a foundation model ground a drone-observed visual instance to its exact real-world entity in a vector map, and reason using map knowledge? Every multiple-choice question pairs a real UAV image with the OpenStreetMap map of the same area, and answering it means matching entities across the two views.

8 tasks at 4 levels 400 questions per task Visual version, north-up OSM raster VETRA, UAV-VisLoc, AirLock, USGS-RSCC
Figure 1: a UAV image and the vector map of the same area, the Visual and Tabular map representations, and the eight tasks at four levels

Swipe sideways to see the whole figure, or tap it to open full size.

Figure 1. UAV-MapVQA pairs UAV imagery with vector maps for instance-to-entity grounding and map-conditioned geospatial reasoning. Each question comes with a rendered raster map (Visual) or a structured entity table (Tabular); this viewer shows the Visual version.

Four levels, eight tasks

Each level adds one capability to the one before it.

Loading questions…
Where the data comes from

Questions are read from all_400.jsonl (400 per task): osm_building_corr/vqa holds T1–T3 and osm_object_context/vqa holds T4–T8, next to their images and debug panels. Your ratings stay in this browser until you export them. To run the viewer locally from the dataset repository: cd ICLR2027-UAV-Dataset && python -m http.server 8000, then open http://localhost:8000/inspect_vqa/inspector.html (add ?file=all.jsonl for the full 500-per-task build).

Build a review set

Choose the tasks and how many questions to review from each.

Tasks
Review status
Questions per task drawn at random from each selected task
0100200300400
20
—
UAV-MapVQA
Test-3.2K quality review
0 of 0 rated
Loading…