Scotopic FOVEA.

It reads a camera feed and returns a track. Detection and tracking of small aerial targets run on the compute next to the sensor — onboard an aircraft or fixed on a mast — and what leaves the node is 26 bytes of track state per update rather than video. Where that track goes, and what your system does with it, is your architecture.

We have not characterised one, and we will not quote one. Range is not a property of the detector — it is a property of the optics in front of it. What we publish is the pixel floor: 8 px minimum resolved size. Put your sensor and lens into the detection geometry calculator and it returns the distance at which your target crosses that floor. A surveyed-truth range trial is on the roadmap and will be published measured, whatever it shows.

Because a range figure is only true for the optics it was measured behind, and that configuration is almost never stated alongside it. Two integrators running the same detector behind different lenses get answers that differ by a factor of three. Publishing the floor lets you compute your own number, and check ours. It is the less flattering way to present the same fact.

Raspberry Pi 5 (8 GB) with a Hailo-8 accelerator, 26 TOPS — roughly €200 of commercial compute per node. It processes the whole 4K field of view at 5.76 FPS, 173.6 ms per frame, measured 13 Aug 2026 on that board. CPU-only, without the accelerator, inference is 595 ms p95 at 1280×1280. No custom silicon, no cloud dependency, no licence server.

No, and that is the point. It sits beside what you already have: it does not touch the autopilot, the navigation or the radio. It consumes a camera feed and emits a track on a versioned contract. Autonomy stacks ask you to adopt their operating system, their board and their qualification path, and you end up reselling someone else’s aircraft. This asks for a camera feed.

Neither. Detection and tracking depend on the sensor and the compute next to it, not on GPS, the datalink or a ground station. Track output is relative — bearing, elevation and size, each carrying its own 1σ uncertainty, so it drops into a fusion filter without being re-derived. World-frame geolocation needs host position and attitude from your platform, and we report the error that adds.

It keeps running at the sensor and the output buffers. On reconnect the buffer flushes losslessly, so you get the interval rather than a gap. At 26 bytes per update — 2.1 kbps sustained at 10 Hz for one track, 7.8 kbps at five — the buffer is small enough that this is a design choice, not a compromise. Against a measured 13.16 Mbps video downlink that is a ratio of 6,327 : 1.

The pipeline is domain-independent; the detector is not. Tracking with coasting through occlusion, the 26-byte contract with decomposed uncertainty, and the quantised edge envelope on Hailo-8 carry over unchanged to other target classes. The detector does not: weights are trained on a specific domain, and retargeting to surface vessels, ground vehicles or people needs labelled data from that domain. We run that as a scoped programme rather than listing it as a capability, and we publish measured performance for small aerial targets only — because that is the only domain we have measured.

57.2% is cross-dataset and zero-shot, on a dataset the detector has never seen. In-distribution, on a held-out split, it is 86.5%. Nearly every figure published in this sector is the second kind, presented without saying which kind it is. We publish the pair because the pair is the informative thing. When you compare vendors, the useful questions are not the number: which split did it come from, what size band, and what denominator did the false-alarm rate use?

No. Access to the detector runs as a scoped evaluation, on your hardware, conducted together. The evaluation methodology and the track contract specification are documents we will send you; the detector is not distributed as a downloadable model file. We would rather say that plainly than publish a download button that turns into an email thread. The geometry calculator is open and needs nothing from you — it will tell you whether the physics works before anyone talks about software.

Not yet, and we will not claim it until it is validated. The interface targets BSI Flex 335 v2.0:2024-03, the current version of the SAPIENT interface control document written by Dstl for the UK MoD. Implementation is planned and has not been validated against a certified decision-making module. The architecture was built for it — local processing, summaries rather than raw feeds — from the first commit.

Nothing. It is a perception layer: it consumes pixels and emits a track with its uncertainty. There is no actuator interface, no flight control and no guidance loop in it — those are different problems, and this module solves one. Where the track goes, and what your system does with it, is yours to design.