Online HD map construction from sensor data, targeting full Lanelet2 fidelity — including traffic signs, traffic lights, bike lanes, and complete topological connectivity. KITScenes Multimodal is the first public dataset to provide all of these elements simultaneously.
Stay tuned for the KITScenes Multimodal Challenges!
Community leaderboard coming soon.
Preview the dataset on HuggingFace ↗Average precision (AP ↑) per map element category. Classes grouped into Lane Markings (LM), Lane Centerlines (LC), Road Infrastructure (RI), Traffic Lights (TL), and Traffic Signs (TS). For MapQR-Topo the topology score (APtopo) is additionally reported.
| Model | APavg ↑ | APLM ↑ | APLC ↑ | APRI ↑ | APTL ↑ | APTS ↑ | APtopo ↑ |
|---|---|---|---|---|---|---|---|
| MapTRv2 | 5.1 | 18.0 | 6.7 | 5.8 | 3.0 | 8.1 | — |
| SDTagNet | 4.5 | 19.4 | 7.1 | 6.3 | 2.4 | 9.0 | — |
| MapQR-Topo | 4.1 | 16.0 | 5.9 | 3.6 | 1.9 | 5.6 | 16.4 |
APtopo reported only for MapQR-Topo, which is the only method designed to predict full topological connectivity. Road Markings (RM) category not reported.
API
Download