TL;DR
Corvus ISR has published a reproducible synthetic benchmark showing that its v2 tracker reduced identity switches by about 42% in baseline and dense tests against its v1 model. The results are company-published and have not been independently replicated in the information available, while both trackers still recorded thousands of errors under stress.
Corvus ISR has published a reproducible synthetic benchmark reporting that its updated multi-object tracker cut identity switches by 42.1% in a 150-mover baseline test and 42.7% in a denser 400-mover test. The results indicate better continuity when following many moving objects, but both tracker versions continued to produce thousands of identity errors under demanding conditions.
In the baseline configuration, which simulated 150 moving objects at two frames per second, identity switches fell from 2,042 to 1,183 per minute. With 400 movers, the reported rate dropped from 14,032 to 8,040 switches per minute. Those changes correspond to reductions of 42.1% and 42.7%, respectively.
Corvus ISR said each comparison used the same fixed-seed synthetic scene, seed 1337, with a 20-second warm-up and 120 seconds of measurement. The sensor model, generated detections and metric definitions were held constant, leaving the tracking model as the tested variable. Because the demonstration is entirely synthetic, it contains no images of real people, vehicles or locations and provides a known identity for every generated object.
The smaller gains recorded in other tests included 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. Corvus also reported that the v2 tracker averaged about 1.2 milliseconds per sensor tick with 400 movers, with its slowest result near five milliseconds against a 10-millisecond browser-processing budget.
Identity Continuity Improves Under Load
Identity switches occur when a tracking system assigns a different track label to the same object across successive frames. Reducing them can make movement histories more coherent and lower the risk that analysts or downstream software confuse one object with another. The reported improvement is most pronounced in the standard and high-density configurations, where reliable association becomes harder as objects move close together.
The public test also gives readers a way to examine failure rates rather than selected successes. Still, the remaining error totals limit what can be inferred about operational readiness. At 400 movers, v2 produced 8,040 switches per minute under the benchmark’s strict measurement rule.
As an affiliate, we earn on qualifying purchases.
V2 Replaces Greedy Association
The archived v1 tracker uses two-pass greedy nearest-neighbour association, constant-velocity prediction and a fixed two-second coasting period. Corvus describes it as a deliberately simple performance floor that remains available in the first two archived demonstration slices.
The v2 model, called “confirmed-track auction,” appears in the third demonstration slice. It adds track confirmation, three-tier auction association, velocity-consistency gating, a noise-scaled reservation price and confidence-decayed coasting. Corvus applies a stricter identity-switch definition than the MOTChallenge IDSW measure: fragmentations, reacquisitions and every change in the identity assigned to a ground-truth object count as switches.
“Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”
— Corvus ISR’s stated publication principle
Real-World Performance Is Untested
It is not yet clear whether the same reductions would appear with real sensor footage, imperfect labels or changing environments. Synthetic ground truth supports exact scoring, but it does not reproduce every source of uncertainty found in live wide-area motion imagery.
The available material does not identify an external institution that replicated the results, provide statistical uncertainty across multiple random seeds or compare v2 with independent commercial and research trackers. 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 separate validation report is identified. The benchmark also holds detections constant, so it does not measure improvements in sensor detection performance.
Future Trackers Face Same Seed
Corvus ISR says each future tracker will be added as a new public row using the same fixed seed. The immediate test will be whether later versions reduce the remaining identity errors while staying within the 10-millisecond processing budget. Broader evaluation across additional seeds and real-world datasets would provide stronger evidence about how well the reported gains transfer beyond this demonstration.
Key Questions
What did Corvus ISR improve?
Corvus ISR changed the software responsible for matching detections to existing tracks. Its v2 confirmed-track auction model recorded 1,183 identity switches per minute in the baseline test, compared with 2,042 for v1, a 42.1% reported reduction.
Were real people or locations used in testing?
No. Corvus ISR describes the product and benchmark as entirely synthetic. Every pixel, moving object and location is generated, providing known ground-truth identities without depicting real people or places.
Did detection accuracy also improve?
No detection improvement was tested. The detection stream was identical for both trackers by design, allowing the comparison to isolate changes in track association. The published reductions concern identity switches, not detection rate.
Can readers reproduce the benchmark?
Corvus ISR says visitors can open its public demonstration and select “Run benchmark” without registering or signing an NDA. Reproduction can confirm how the test behaves in the published browser demonstration, but independent testing on other datasets would be needed to judge wider performance.
Source: Thorsten Meyer AI
Source: Thorsten Meyer AI
