Distilling Vision Models 100× Smaller Than Foundation Backbones

Large vision foundation models are powerful teachers — but they are often the wrong artifact to deploy on a forecourt NVR or browser tab.

The deployment question first

Before choosing an architecture, answer four constraints:

1. What event must an operator act on?
2. What latency and hardware budget are real?
3. What false-positive rate will make people ignore alerts?
4. How will the model be updated, rolled back, and audited?

Those answers usually eliminate multi-gigabyte backbones long before accuracy discussions begin.

Our distillation approach

We distill teacher ensembles into YOLO-class students with:

• Task-specific heads for person, vehicle, and fire skills
• Hard-negative mining from operational scenes
• ONNX export paths validated on CPU and embedded targets
• Regression clips that catch scene-specific regressions after each export

Distillation is an engineering tradeoff: retain the behavior that supports the target task, remove capacity that cannot be used at the edge, and prove the result on the intended runtime.

Why the runtime changes the model

The size of a checkpoint is not the same as the cost of a deployment. Runtime choice, preprocessing, memory pressure, batch size, thermal limits, and update mechanics all change what a usable detector looks like. We treat export and inference validation as part of model development, rather than a handoff at the end of research.

A smaller artifact also makes versioning, field testing, and rollback materially easier for the teams operating it.

How to read size and accuracy claims

On person detection, our distilled Sentinel student reaches 92% accuracy under Innomium crowd protocols at a footprint dramatically smaller than general-purpose segmentation backbones. The “100× smaller” framing compares against heavy foundation-style teachers — not against every competing edge detector, and not as a universal guarantee on your cameras.

Always re-measure:

• precision/recall on your held-out scenes
• latency on the target device at the intended resolution
• false positives that matter to operators
• end-to-end time from frame to actionable event

A distillation checklist for production teams

• Freeze the operating definition of a true positive before tuning thresholds.
• Keep a hard-negative library that includes the site’s known distractors.
• Validate the exported ONNX path, not only the training framework checkpoint.
• Record memory, thermal, and sustained FPS under realistic concurrency.
• Version the preprocessing contract with the weights.
• Decide who owns threshold changes after go-live.

When not to distill

If the problem still needs open-vocabulary understanding, frequent class expansion without data, or research exploration rather than a fixed mission skill, a larger model or a different pipeline may be the better first step. Distillation shines when the skill is narrow, the hardware is constrained, and the evaluation set is honest.

Public artifacts: Sentinel, Vantage, and Ember on Hugging Face. Method context: [vision layer](/updates/building-the-innomium-vision-layer) and [evaluation protocol](/updates/edge-vision-evaluation-protocol).