Research map / Detection and segmentation / Open-set and few-shot detection
OpenSatMap: A Fine-grained High-resolution Satellite Dataset for Large-scale Map Construction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces OpenSatMap, a large fine-grained, high-resolution (level-19/20) satellite dataset with instance-level vectorized lane/curb annotations aligned with nuScenes and Argoverse 2, plus a baseline line-detection pipeline and a satellite-enhanced online map construction experiment.
Problem
Existing satellite road datasets offer only coarse semantic labels, low resolution, small scale, and no alignment with autonomous driving benchmarks, which hinders fine-grained large-scale map construction from satellite imagery.
Approach
A manually annotated satellite dataset with instance-level polyline lane annotations (3 categories, 8 attributes) at level-20 (0.15 m/pixel) and level-19 resolution, geographically diverse and spatially aligned with nuScenes and Argoverse 2, paired with a simple segmentation-then-vectorization baseline.
Key results
- On OpenSatMap20 the baseline reaches 33.69 mIoU and 20.30 Chamfer AP at 0.9 m threshold, versus 28.71 mIoU and 16.04 on OpenSatMap19.
- SatforHDMap with satellite input reaches 50.9 IoU overall on nuScenes semantic map segmentation, versus 32.9 for HDMapNet.
- The dataset contains 446,645 line instances across 3,787 images (1806 at 2048x2048 and 1981 at 4096x4096) from 60 cities in 19 countries.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add corresponding satellite imagery | HDMapNet for online map construction · nuScenes validation | Divider IoU: 40.6 -> 50.2 | Helped |
| Add corresponding satellite imagery | HDMapNet for online map construction · nuScenes validation | Crossing IoU: 18.7 -> 53.2 | Helped |
| Add corresponding satellite imagery | HDMapNet for online map construction · nuScenes validation | Boundary IoU: 39.5 -> 49.4 | Helped |
| Add corresponding satellite imagery | HDMapNet for online map construction · nuScenes validation | All IoU: 32.9 -> 50.9 | Helped |
Surprising findings
- Instance-level AP scores remain low even when qualitative predictions look visually satisfactory, which the authors attribute to strict matching criteria and segmentation errors at attribute changes and adjacent lines.
Limitations
- Imagery from Google Maps is not updated in real time and may be outdated relative to nuScenes and Argoverse 2 capture times.
- Level-20 imagery is unavailable or cloud-covered in some regions, limiting geographic coverage.
- Annotation subjectivity across about 60 annotators can cause inconsistent labels.
Experimental setup
Datasets: OpenSatMap nuScenes Argoverse 2
Benchmarks: OpenSatMap nuScenes
Metrics: Chamfer AP mask AP mIoU IoU
Baselines: HDMapNet SatforHDMap
Builds on: SegNeXt SatforHDMap HDMapNet
Closest papers by meaning
- SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.87)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.84)
- Patched Line Segment Learning for Vector Road Mapping (AAAI 2024 · similarity 0.84)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.83)
- GeoLink: Empowering Remote Sensing Foundation Model with OpenStreetMap Data (NeurIPS 2025 · similarity 0.83)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.