SpaceNet
SpaceNet
Labeled satellite imagery for building footprints and roads, built to give computer vision a public benchmark.
An open-innovation repository launched in 2016 that hosts precision-labeled high-resolution satellite imagery with co-registered map features such as buildings and roads. It also hosts datasets from other projects such as IARPA's Functional Map of the World. Used for challenges and model training.
Access
- Cost
- FreeNo charge to download; hosted in an AWS S3 bucket.
- License
- Provider's own termsVarious licenses by dataset (see SpaceNet datasets page)
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadsee documentation
What's in it
- One row is
- One image tile with its co-registered labels
- Covers
- Selected cities and areas of interest worldwide
- Years
- 2016-present
- Updated
- Quarterly
- Size
- Not stated
Quality
Long-running, well-known benchmark with clearly labeled tasks. Each sub-dataset has its own terms and layout.
- License varies by sub-dataset
- Coverage limited to chosen areas of interest
Connects to
- RarePlanes: Another labeled satellite imagery set, focused on aircraft
- OpenStreetMap: Source of the kind of map features SpaceNet labels are compared with
Questions it could answer
- How well do building detectors trained in one city transfer to another?
- Which road types are hardest to extract from imagery?