Vantage is our vehicle-detection skill for roadways, parking structures, and logistics yards. It prioritizes reliable boxes on partial occlusions and night glare — the cases that break general-purpose detectors.
Specs
- 93% accuracy on Innomium highway validation splits
- Sub-20MB ONNX with documented preprocessing
- Live Space demo for side-by-side evaluation
Read the percentage as a protocol snapshot. Night glare, rain, headlight bloom, partial occlusion, and tight camera angles can distort a model that looks strong on generic data.
From detection to traffic intelligence
Vehicle detection becomes useful when it is connected to a decision: occupancy, queue length, access control, incident review, or a traffic-flow measure. We design deployments around that downstream outcome, including the camera coverage, event schema, confidence rules, and retention policy required by the wider system.
Scene-specific validation and replayable regression clips make it possible to improve the detector without losing track of the operating conditions that matter.
Open evaluation path
Use the Hugging Face demo for a first impression, then rebuild the metric on your own cameras. If a challenge round is open on Arena, treat leaderboard results as research evidence under challenge rules — not as a substitute for site acceptance testing.
Related: [distilling vision models for the edge](/updates/distilling-vision-models-for-the-edge).