CORVUS ISR’s Auction Tracker Cuts ID Switches by More Than 42%

AIThis post was created with the assistance of artificial intelligence (AI).

Corvus ISR tracker benchmark matrix (seed 1337) — The published matrix — every row reproducible. Source: corvusisr.com/benchmark

CORVUS ISR is a wide-area motion imagery (WAMI) exploitation product with the public benchmark comparing two tracker models on an identical fixed-seed synthetic scene. Each row uses seed 1337, a 20s warm-up and 120s measured period, with perfect ground truth.

Multi-object tracking connects detections across frames while trying to preserve each object’s identity. CORVUS ISR counts EVERY change of the track identity assigned to a ground-truth object, making its ID-switch metric deliberately stricter than the MOT-challenge IDSW definition: fragmentations and re-acquisitions count as switches too.

The v1 “greedy nearest-neighbour” baseline uses two-pass greedy association, constant-velocity prediction and fixed 2s coasting. It is deliberately simple and remains the published floor, runnable in archived demo slices 1-2. The current v2 “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.

With the sensor model, detection generation and metric definitions byte-identical—only the tracker differs—the baseline configuration of 150 movers at 2fps falls from 2,042 to 1,183 ID switches per minute (−42.1%). At 400 movers, the dense configuration drops from 14,032 to 8,040 (−42.7%). Detection rate is a sensor property and is identical for both models by construction.

Corvus ISR live demo — The live demo — press “Run benchmark” to reproduce the numbers. Source: corvusisr.com/demo

The gains remain measurable under other synthetic stresses: the frame-starved 0.5fps row improves by −16.6%, the occluded 20% row by −18.6%, and the degraded 1fps + jitter, 70% contrast row by −18.1%.

At density 400, v2 averages ~1.2ms per sensor tick, with a worst result of ~5ms against a 10ms budget, so it runs in real time in the browser. Anyone can reproduce it live by opening the demo and pressing “Run benchmark”—no signup and no NDA.

Both models still commit thousands of identity errors per minute under stress, and CORVUS ISR deliberately publishes those rows on its own site. Synthetic scenes provide perfect ground truth, making these measurements, not marketing, and every future tracker must appear as a new public row against the same seed. “Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”

V2 was built by an AI executor against a written acceptance contract and independently reviewed before shipping. Everything is fully synthetic—there are no real persons, vehicles or places, and every pixel is generated—so readers are invited to run the benchmark themselves.