EU Scenario Miner

Finding European road scenarios (cobblestones, zebra crossings, hatched no-drive areas, roadworks barriers, speed bumps, electronic signs, tractors) in the A2D2 driving data from southern Germany. A C++ scanner reads the semantic labels and gives each frame its scenario tags. Plain-English CLIP text queries then search the camera frames without labels, and every query is scored against those tags. Code: github.com/pavanyadava007/eu-scenario-miner.

Try a query

Preset queries run at once. A free-text query first loads the CLIP text encoder (127 MB, fp16, cached by the browser afterwards).

Scenario tags from the labels

How well do text queries find them?

C++ label scanner vs numpy

Numbers come from results/results.json as written by scripts/run_all.py. A2D2 is CC BY-ND 4.0 by Audi AG; this page ships no A2D2 images, only frame ids, tags and image embeddings. Author: Pavan Yadav Annappa.