Corvus ISR’s Public Experiment Highlights Major Drop In Tracker ID Switches

Corvus ISR has published a reproducible synthetic benchmark reporting that its updated multi-object tracker produced about 42% fewer identity switches than its earlier baseline in standard and dense traffic configurations. The reduction could improve object identity continuity in wide-area motion imagery, although the results come from the project’s own synthetic test environment and have not been independently replicated in the material provided.

In the benchmark’s baseline configuration of 150 moving objects at two frames per second, reported identity switches fell from 2,042 to 1,183 per minute, a 42.1% reduction. With 400 moving objects, the rate dropped from 14,032 to 8,040 per minute, or 42.7%, according to the published matrix.

Smaller gains appeared in other stress tests. Corvus ISR reported 16.6% fewer switches at 0.5 frames per second, 18.6% fewer with 20% occlusion and 18.1% fewer in a degraded test combining one frame per second, jitter and 70% contrast. Detection rates were held equal between the two models because the benchmark changes only the tracker, not the simulated sensor or detections.

Every row uses the same fixed-seed synthetic scene, identified as seed 1337, with a 20-second warm-up and 120-second measurement period. Corvus ISR says the sensor model, generated detections and metric definitions remain identical across runs. The environment contains no real people, vehicles or locations, allowing the system to retain exact ground-truth identities for every simulated object.

Identity Continuity Improves Under Density

Multi-object tracking depends on preserving an object’s assigned identity from one frame to the next. Fewer identity switches mean analysts or downstream software are less likely to confuse one vehicle with another, particularly when many objects move close together or disappear briefly from view.

The dense-scene result also addresses processing speed. At 400 objects, Corvus ISR reports that v2 averaged about 1.2 milliseconds per sensor tick and reached roughly five milliseconds in its slowest run, below a 10-millisecond processing budget. Those figures indicate browser-based real-time operation in this test, but they do not establish performance on live imagery or other hardware.

Auction Model Replaces Greedy Baseline

The archived v1 tracker uses greedy nearest-neighbour association, constant-velocity prediction and fixed two-second coasting. Corvus ISR describes it as a deliberately simple published performance floor that remains available in the first two demo slices.

The v2 model, called confirmed-track auction, appears in the third slice. It adds track confirmation, three-tier auction association, velocity-consistency gating, a noise-scaled reservation price and confidence-decayed coasting. Thorsten Meyer AI says the tracker was built by an AI executor against a written acceptance contract and reviewed independently before release, but no reviewer or review report was identified.

“Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”

— CORVUS ISR publication principle

Live-Imagery Performance Is Untested

It is not yet clear whether the reported gains would carry over to real sensor footage, different motion patterns, longer runs or scenes outside the fixed seed. Synthetic ground truth makes precise measurement possible, but it cannot capture every source of noise, occlusion and detection error found in operational imagery.

The benchmark also uses a stricter identity-switch definition than the MOTChallenge IDSW measure, counting fragmentations and reacquisitions as switches. That choice helps explain the large totals, but it limits direct comparison with outside benchmarks. Both trackers still recorded thousands of switches per minute in demanding rows, and the supplied material includes no third-party replication, statistical uncertainty range or comparison with other tracking systems.

Future Trackers Face the Same Seed

Corvus ISR says each future tracker version will be added as a new public benchmark row using the same fixed seed and measurement rules. The next test will be whether outside users can reproduce the published numbers and whether later experiments add independent validation, more varied scenes or real-world imagery.

Source: Thorsten Meyer AI

Key Questions

What changed in the Corvus ISR tracker?

The project replaced a greedy nearest-neighbour baseline with a confirmed-track auction model that adds confirmation rules, tiered association, velocity gating and confidence-based coasting.

How large was the reported improvement?

Identity switches fell by 42.1% with 150 movers and 42.7% with 400 movers. Other stress configurations produced reductions ranging from 16.6% to 18.6%.

Does the benchmark use real surveillance footage?

No. It uses a fully synthetic scene containing no real people, vehicles or places. That design supplies exact ground truth but leaves performance on live footage unresolved.

Were the results independently verified?

No independent reproduction was included. Thorsten Meyer AI says the tracker received an independent review before release, but that claim is separate from third-party benchmark validation.

Can the public rerun the experiment?

Corvus ISR says users can select “Run benchmark” in its public demo without an account or NDA. The fixed seed and published settings are intended to make each benchmark row reproducible.

Source: Thorsten Meyer AI