- Home
- Technology
Stack
The vision stack
Four layers between a photon hitting a sensor and a decision a professional can defend. Here is what happens in each of them.
Layers
Four layers, photon to decision.
Sensing — choosing what the physics will actually give you
No single band sees everything. Low-light colour gives classification detail and fails in true darkness. Uncooled LWIR sees thermal contrast through darkness and light haze but carries little detail. Cooled MWIR reaches far but costs and needs cooldown. SWIR cuts haze that defeats both. We specify a band mix against the operating profile rather than defaulting to whatever is cheapest, and we calibrate the whole rig as one rigid assembly so the geometry is trustworthy.
Perception — detection, classification and tracking on the stitched scene
Detection runs on a stitched panorama rather than per-camera, which removes the seam artefacts that fragment a track into two. Classification covers fourteen marine classes weighted toward the non-cooperative traffic that matters in Indian waters. Tracking is multi-hypothesis with re-identification, so a contact that disappears behind swell for four seconds comes back as the same track rather than a new one.
Fusion — one picture from cameras, radar, AIS and the chart
Vision contributes what radar and AIS cannot: vessel type, activity, light and shape patterns, and the small non-metallic craft that never generate a usable return. Radar contributes range accuracy at distance. AIS contributes identity for those who transmit. The fusion layer associates them into single tracks and — importantly — records which sensor contributed what, so a track's provenance is inspectable.
Decision — rule-cited reasoning, logged
For autonomy, the COLREG rule engine produces manoeuvres with the rule that generated them, the give-way determination and the CPA being protected. For inspection, findings arrive with a measurement, a position and a confidence band, queued for human confirmation. In both cases the output is a defensible recommendation rather than an unexplained instruction.
Training data
Models inherit the sea they were trained on.
This is the single most important technical claim we make, so it is worth being precise about it.
A detector's performance is bounded by the distribution of its training set. A model trained predominantly on North Sea and Mediterranean footage has seen a particular catalogue of hull forms, a particular sea-state distribution, a particular set of atmospheric conditions and a particular density of AIS-transmitting traffic. Deploy it off Kochi in August and every one of those priors is wrong at once.
We collect and annotate in Indian waters across the seasonal cycle, with deliberate weighting toward the conditions where imported systems degrade: south-west monsoon rain, post-monsoon haze, dusk and dawn glare on a flat sea, and dense unlit small-craft traffic.
| Band | Strength | Limit |
|---|---|---|
| Low-light colour | Classification detail, light and shape reading, hull markings | Fails below starlight; defeated by glare |
| Uncooled LWIR | Thermal contrast in full darkness; person-in-water detection | Low detail; degraded by heavy rain |
| Cooled MWIR | Long-range detection and recognition | Cost, cooldown time, import dependency |
| SWIR | Haze and thin-fog penetration | Needs illumination at night; costly |
| Radar (fused) | Range accuracy, all-weather at distance | Sea clutter buries small wooden craft |
| AIS (fused) | Identity and intent for cooperative traffic | Absent on most small craft; can be switched off |
Deployment
On the hull, not in a datacentre.
Bandwidth over the Indian Ocean is expensive, intermittent, and for naval users frequently unavailable on purpose. Anything safety-relevant runs on the vessel.
- Local inference by default. Every safety-relevant decision is made on SETU-EDGE aboard.
- Opportunistic sync. Shore synchronisation is store-and-forward; a lost link degrades reporting, never safety.
- Signed model artefacts. The node rejects unsigned models, so what runs in service is what passed trials.
- Air-gap capable. SAMUDRA-C2 deploys into a fully disconnected enclave when the customer requires it.
- India-resident data. Nothing leaves the country unless a customer explicitly configures it to.
What the stack does not do.
Vision is not a replacement for radar, and we will not sell it as one. In heavy rain and thick fog, optical performance degrades in ways radar does not. At long range, radar's accuracy is better than horizon-geometry estimation.
Vision earns its place where radar and AIS are weak: close in, against small non-metallic craft, for vessel-type and activity determination, and for anything requiring passive operation. A well-specified system uses all three.
Start a conversation
Want the technical detail?
Aparna and the engineering team will walk through interface control documents, model performance methodology and trials design with your technical staff.