emb-ai/traffic-sign-bench

Traffic Sign Bench

Official per-sign SUMO maps for TrafficSignBench: real Moscow OSM layouts, 25 signs, 2500 maps. Protocol size is 80 train + 20 test maps per sign.

Road geometry is derived from OpenStreetMap © OpenStreetMap contributors and is released under ODbL 1.0.

Examples

Roundabout

Curved segment

Straight segment

Junction

Download

All scenes land under data/scenes/<sign>/<scene_id>/, which is what eval expects:

huggingface-cli download emb-ai/traffic-sign-bench \
    --repo-type dataset \
    --local-dir data

Code and runners: https://github.com/emb-ai/traffic-sign-bench.

from datasets import load_dataset
ds = load_dataset("emb-ai/traffic-sign-bench", split="train")

That loads the catalog (one row per scene, with a path to the preview and the net). The .net.xml files themselves are next to the catalog in scenes/.

Layout

scenes/<sign>/<scene_id>/
  map.net.xml
  meta.json
  custom_cropped.png
metadata/
  signs.yaml
  sign_allocations.json
  train_ids.json
  test_ids.json
  catalog.jsonl
  catalog.parquet

sign is the eval id (folder name under scenes/). Train/test is in sign_allocations.json and the catalog, not in the folder name.

Signs

Sign	Family	Train	Test	Total
blocked_road	junction	79	19	100
main	junction	73	20	100
roundabout	junction	38	11	100
secondary	junction	80	18	100
stop	junction	75	19	100
yield	junction	77	20	100
direction_left	dual_path	80	20	100
direction_left_right	dual_path	80	20	100
direction_right	dual_path	80	20	100
direction_straight	dual_path	80	21	101
direction_straight_left	dual_path	80	19	100
direction_straight_right	dual_path	80	20	100
no_entry	dual_path	57	14	100
no_turn_left	dual_path	80	20	100
no_turn_right	dual_path	80	20	100
one_way_left	dual_path	80	20	100
one_way_right	dual_path	80	19	99
crosswalk	segment	80	20	100
detour_either	segment	80	20	100
detour_left	segment	80	20	100
detour_right	segment	80	20	100
min_speed	segment	80	20	100
residential_zone	segment	80	20	100
speed_limit	segment	80	20	100
zone_speed_limit	segment	80	20	100

How it was built

Moscow OSM → city SUMO net → junction / dual-path / segment crops → each sign queries that shared pool (signs.yaml) → materialize into data/scenes/<sign>/. Pedestrian-crossing scenes get a mid-block zebra injected after copy.

Split is stamped on place identity before allocation:

• junctions and dual-path: junction_id
• segments: osm_way_id (one street does not appear in both halves)

Signs may reuse the same physical place within a split when allowed by the tiered assign policy (same behavioral family, or same semantic group).

Limitations

• Geography is Moscow only.
• On some pedestrian-crossing segments SUMO omits the crossing edge; the split node and sidewalks are still there.
• These folders are maps, not closed-loop eval manifests.

Citation

@misc{traffic-sign-bench,
  title  = {Traffic Sign Bench},
  author = {EMB AI},
  year   = {2026},
  url    = {https://huggingface.co/datasets/emb-ai/traffic-sign-bench}
}