Sentinel is Innomium's person-detection skill for mission environments where generic models fail — overlapping crowds, variable lighting, and tight latency budgets.
What changed
We combined test-time augmentation, adaptive confidence calibration, and scene-aware distillation to lift accuracy on held-out crowd splits while keeping the model at 19MB ONNX.
How to read the 92% number
The published accuracy reflects an Innomium evaluation protocol on defined crowd splits. It is not a guarantee for every airport, stadium, or retail floor. Lighting, camera elevation, occlusion density, and class definitions all move the operating point. Treat the number as a starting evidence snapshot, then rebuild the metric on your own held-out clips.
Where teams evaluate it
- Airport concourses and security lanes
- Stadium ingress and egress monitoring
- Dense retail floors during peak hours
How to evaluate a crowd detector
Accuracy is only one part of a useful operating picture. Dense scenes need a review of missed detections at entrances, false positives caused by signage or reflections, and performance as lighting and camera elevation change. A deployment plan should set a confidence threshold for each scene, capture representative clips for regression testing, and measure the time from frame to usable event.
For safety and operations teams, the right question is not simply whether a box appears on screen. It is whether the detector produces a dependable signal inside the existing video-management and incident-response workflow. That is why we test the model together with the export format, event rules, and integration surface.
Try it today
The live Hugging Face Space runs entirely in-browser. For integration experiments, pull weights from the published model card and validate against your scenes before production use. Commercial terms follow the license on the model card.
See also: [Building the Innomium Vision Layer](/updates/building-the-innomium-vision-layer).