How it works
We break your perception model on purpose — then show you exactly where, and fix it.
Siltframe takes frames from your vehicle, renders the conditions your model hasn’t seen (dust, night, rain, mud, fog) with physics rather than filters, and measures what happens. You get two things: a failure map of your model, and labelled hard-case data that wins the loss back.
Your frames
A few hundred labelled camera frames from your platform — or 3–5 site photos for a first look. Your model as PyTorch or ONNX if you want a failure map.
Depth per pixel
A monocular depth model estimates how far every pixel is. Haze and dust depend on distance, so this is what makes them physically right.
Physics-based degradation
Dust, night, rain on the lens, mud on the lens and fog, each at three severities. Your labels carry over unchanged — the ground is still where it was.
Output A · Failure map
How much, where, and on which classes your model breaks
- Your model scored on your frames under every condition and severity
- Checked with independent (held-out) generators too, so the result isn’t an artefact of our own simulator
- Per-class and drivable-surface IoU, a failure gallery and a ranked fix list
Output B · Hard-case data
Labelled frames that win the loss back
- Rendered on your own scenes, in your label format, with a commercial licence
- Measured against a control fine-tune on clean data, over several seeds
- Reported on your real held-out frames, so you see the gain that survives contact with reality
What it did on public data
SegFormer-B0 trained on clear RELLIS-3D frames, then tested on degradations it never saw:
How an engagement runs
- Brief (day 0). You send photos or frames and the conditions you care about. We sign an NDA if you need one.
- Preview (≤ 2 business days). Sample frames rendered on your own scenes and a fixed quote.
- Failure map (≈ 10 business days). Your model under every condition, with a ranked list of what to fix.
- Hard-case pack. Labelled frames for the conditions that hurt most, plus a before/after on your real validation set.
what we won’t claim
- Synthetic data doesn’t replace real adverse-weather testing. It finds weaknesses early and cheaply; your real validation frames confirm the fix.
- Gains measured on our own simulator are larger than on independent ones — so we always report both, and lead with the smaller number.
- Your data is used only for your project and deleted on request. See privacy.