Snapshot Serengeti camera trap images
Snapshot Safari / University of Minnesota, hosted by LILA BC
7.1 million camera trap photos from the Serengeti, labeled by volunteers, and about 76% of them empty.
Snapshot Serengeti ran a grid of motion-triggered cameras in Serengeti National Park, Tanzania, and volunteers on Zooniverse labeled the images. The LILA release has about 2.65 million image sequences (7.1 million images) from seasons one through eleven, with labels in 61 categories, mostly species, plus about 150,000 bounding boxes. It is a benchmark for wildlife detection AI.
Access
- Cost
- FreeFree download from public cloud storage, no account.
- License
- Open licenseCommunity Data License Agreement, permissive variant
- Sign-up
- No sign-up
- How you get it
- Bulk downloadJPEG, JSON (COCO Camera Traps), CSV
What's in it
- One row is
- One camera trap image sequence with species labels
- Covers
- Serengeti National Park, Tanzania
- Years
- Seasons 1-11 (2010s)
- Updated
- One-time release
- Size
- About 7.1 million images in 2.65 million sequences
Quality
Large, well-documented, and published in Scientific Data. Volunteer consensus labels are good but imperfect for rare and look-alike species.
- About 76% of images are empty
- Rare species have few examples
- Labels are crowd consensus
Connects to
- iNaturalist Licensed Observation Images (Open Data on AWS): Broader labeled wildlife photos for model training
- Movebank animal tracking database: GPS tracks of Serengeti species for comparison
Questions it could answer
- How do lions and hyenas shift their activity by time of day?
- How well do detection models trained here transfer to other parks?