TL;DR
Corvus ISR has published a reproducible synthetic benchmark reporting that its v2 tracker reduced identity switches by about 42% in baseline and dense tests compared with its v1 tracker. The test uses fixed inputs and perfect ground truth, but the results remain company-published and both trackers record thousands of errors under heavy stress.
Corvus ISR has published a public synthetic benchmark reporting that its current v2 tracker cut identity switches by about 42% in baseline and dense tests compared with its earlier v1 model. The result matters because preserving object identity across video frames is a core test of multi-object tracking, although the figures were published by the product’s developer and no outside replication was cited.
In the test with 150 moving objects at two frames per second, identity switches fell from 2,042 to 1,183 per minute, a reported reduction of 42.1%. In the denser test with 400 moving objects, the rate dropped from 14,032 to 8,040 switches per minute, or 42.7%.
The benchmark also reports smaller gains under tougher conditions. V2 recorded 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 constant because detection generation is identical for both trackers by design.
Corvus ISR says each row uses the same fixed-seed synthetic scene, seed 1337, with a 20-second warm-up and 120 seconds of measurement. The company says the sensor model, detections and metric definitions remain unchanged, leaving the tracker as the only variable. Users can run the published tests in a browser without registration or a nondisclosure agreement.
Identity Stability Improves Under Load
The reported reductions indicate that v2 is better at maintaining assigned identities through successive frames, especially when many objects are moving at once. Fewer identity changes can make tracked paths more coherent and reduce errors in systems that depend on consistent object histories.
The results also show the limits of the improvement. Even after the reported gains, v2 produced 1,183 switches per minute in the baseline test and 8,040 per minute with 400 movers. Those totals mean the tracker remains error-prone under the benchmark’s strict scoring method and should not be described as eliminating identity loss.
As an affiliate, we earn on qualifying purchases.
V2 Replaces Greedy Association
Corvus ISR is a wide-area motion imagery exploitation demonstration built entirely from generated imagery. The company says no real people, vehicles or locations appear in the product, allowing every simulated object’s correct identity to be known throughout a test.
The archived v1 tracker uses two-pass greedy nearest-neighbour association, constant-velocity prediction and fixed two-second coasting. V2 adds track confirmation, three-tier auction association, velocity-consistency gating, a noise-scaled reservation price and confidence-decayed coasting. Corvus ISR describes v1 as a deliberately simple published floor rather than a competitive model.
The benchmark applies a stricter identity-switch definition than the MOTChallenge IDSW measure. It counts every change in the track identity assigned to a ground-truth object, including fragmentations and reacquisitions, which helps explain the high error totals and limits direct comparisons with results scored under other rules.
“Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”
— Corvus ISR publication principle
Independent Replication Is Still Missing
It is not clear whether an independent party has reproduced the full matrix or audited the implementation behind the browser demo. Corvus ISR says the tracker was independently reviewed before release, but the provided information does not identify the reviewer, describe the review method or publish a separate validation report.
The benchmark also does not establish how v2 would perform on real sensor footage, unfamiliar synthetic scenes or against unrelated tracking systems. Using one fixed seed supports repeatability between the two Corvus models, but it does not by itself show general performance across varied data.
Future Trackers Face Same Seed
Corvus ISR says each future tracker will be added as a new public row against the same seed, preserving the current comparison setup. Readers can check the present claims by opening the demo and running the benchmark, while outside replication and tests across additional scenes would provide stronger evidence of broader performance.
Key Questions
What did the Corvus ISR benchmark find?
It reported that v2 reduced identity switches by 42.1% in the 150-mover baseline test and by 42.7% in the 400-mover test compared with v1.
Does the benchmark use real surveillance footage?
No. Corvus ISR says the demonstration uses fully synthetic imagery containing no real people, vehicles or places. Generated scenes provide exact ground-truth identities for scoring.
Can the results be reproduced publicly?
Corvus ISR provides a browser-based benchmark that it says runs the fixed scene without signup or an NDA. No documented independent replication was identified in the supplied information.
Does v2 run in real time?
Corvus ISR reports an average of about 1.2 milliseconds per sensor tick at 400-object density, with a worst result near five milliseconds against a 10-millisecond browser budget. The platform tested was the browser demo; hardware and browser details were not provided.
Source: Thorsten Meyer AI
Source: Thorsten Meyer AI
