Give the sky your attention.
A local camera captures what passes through its view. Detection starts at the station, close to the source.
Intelligent Anomaly Detection Network
Star People is an observation network for people with wonder.
Join the network ↗See how it works ↓Anomalous
Independently observing.Collectively learning.Always looking up. ✳
01 / A bigger picture
Your patch of sky.
Part of something bigger.
One camera has a point of view. A network has perspective. The idea is simple: independent stations, connected by a shared way of spotting the unexpected.
EXPLORE THE CONCEPTA rooftop.
A rooftop.
A new perspective.
A permanently mounted camera watches its local sky. A small computer looks for motion and classifies what appears.
Shared anomaly model
Model updates
RooftopDetect locally
BackyardDetect locally
Remote siteDetect locally
Detections
Shared observationsPeople review the evidence
Reviewed examples improve the shared model
02 / Curiosity, with a process
Watch. Notice.
Learn together.
Automation does the watching.
People help make sense of it.
Every useful review makes the next question better.
Find the moments that matter.
The shared model evaluates detections and flags unusual activity for a closer look. A flag is a question, not a conclusion.
Turn a sighting into a shared insight.
People review the evidence. Reviewed examples inform training, helping the network learn what deserves attention.
Learning in the open
Different approaches.
Shared evidence.
There’s more than one way to spot the unexpected. We compare human-curated, hybrid, and machine-trained models—and experiment with their decision settings—to find what catches meaningful detections with fewer false alarms.
Human review helps us measure what works. We track accuracy, precision, recall, and false positives, including how results differ between day and night.
Explore the modelModel leaderboard
Comparison snapshot| Approach | Score iA score out of 100 using a weighted harmonic mean: recall 35%, precision 25%, specificity (avoiding false positives) 25%, and accuracy 15%. Higher is better. | Recall | False positives |
|---|---|---|---|
| 01Human-curatedExperimental settings | 87.2 | 92.8% | 0.5% |
| 02Human-curatedStable settings | 84.6 | 95.2% | 0.7% |
| 03HybridHuman + machine labels | 64.4 | 81.9% | 1.7% |
| 04Machine-onlyMachine-generated labels | 23.1 | 43.9% | 5.6% |
105,604 reviewed examples
Training-set comparison, not a measure of performance on unseen footage.
An open record
Public Anomaly Archive
When an observer reviews a detection as anomalous, Field Station automatically prepares and uploads its evidence bundle to the Internet Archive. Manual recordings can be published there, too.
Explore the original video, annotated frames, and accompanying observation data. Reviewed annotations travel with the recording, giving others the context to take a closer look.
Browse the archiveFrom the Field Station archive · Video excerptExplore this observation
About the Project
The ones who stop mid-conversation to look up. Who keep a camera by the window. Who would rather ask a good question than settle for an easy answer.
Our name is a nod to the people searching the skies—and the “star people” imagined in stories of visitors from beyond them. We bring the curiosity. The evidence gets the last word.
There’s room for you here ↗A shared curiosity is a good start.
The sky is big.
Let’s look together.
Built on the Field Station platform and its anomaly vision model. For the makers, the observers, and everyone who wants a closer look.
Explore Field Station ↗Explore the shared model ↗