The published matrix — every row reproducible. Source: corvusisr.com/benchmark
CORVUS ISR is a wide-area motion imagery (WAMI) exploitation product centered on multi-object tracking: maintaining a distinct identity for each moving object across successive frames. This is a fully synthetic demonstration product involving no real persons, vehicles, or places; every pixel is generated.
Its the public benchmark compares two tracker models on an identical fixed-seed synthetic scene with perfect ground truth. Every row uses seed 1337, with a 20s warm-up followed by 120s measured. The sensor model, detection generation, and metric definitions are byte-identical; only the tracker differs.
The v1 “greedy nearest-neighbour” baseline uses two-pass greedy association, constant-velocity prediction, and fixed 2s coasting. It is deliberately simple, represents the published floor, and remains runnable in archived demo slices 1-2. Current v2, the “confirmed-track auction” model in demo slice 3, adds track confirmation, three-tier auction association, velocity-consistency gating, a noise-scaled reservation price, and confidence-decayed coasting.
The live demo — press “Run benchmark” to reproduce the numbers. Source: corvusisr.com/demo
At the baseline configuration of 150 movers and 2fps, ID switches per minute fell from 2,042 to 1,183, a −42.1% change. In the dense configuration with 400 movers, the count dropped from 14,032 to 8,040, or −42.7%. Detection rate is a sensor property and is identical for both models by construction.
The other stress rows show smaller but still measured reductions: −16.6% in the frame-starved 0.5fps case, −18.6% with 20% occlusion, and −18.1% in the degraded 1fps plus jitter, 70% contrast configuration. These results isolate tracker behavior because every other benchmark component remains unchanged.
The identity-switch metric counts every change of the track identity assigned to a ground-truth object. That makes it deliberately stricter than the MOT-challenge IDSW definition because fragmentations and re-acquisitions also count. Both models still commit thousands of identity errors per minute under stress, and CORVUS ISR intentionally publishes those rows on its own site: with synthetic scenes and perfect ground truth, these are measurements, not marketing. “Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”
At density 400, v2 averages ~1.2ms per sensor tick, with a worst result of ~5ms against a 10ms budget, enabling real-time execution in the browser. It was built by an AI executor against a written acceptance contract and independently reviewed before shipping, while every future tracker must appear as a new public row against the same seed. With no signup or NDA required, readers can reproduce it live, press “Run benchmark,” and run the benchmark themselves.