TL;DR
CORVUS ISR published a reproducible synthetic benchmark reporting that its v2 tracker reduced identity switches by 42.1% with 150 movers and 42.7% with 400 movers. The test uses fixed inputs and perfect ground truth, but it is company-hosted and has not been independently replicated.
CORVUS ISR has published a reproducible synthetic benchmark reporting that its current v2 tracker cut identity switches by 42.1% in a 150-object test and 42.7% in a denser 400-object test. The result matters because maintaining the same identity for an object across frames is a central challenge in multi-object tracking, although the company-hosted findings have not yet been independently replicated.
In the baseline configuration, which simulated 150 movers at two frames per second, identity switches fell from 2,042 to 1,183 per minute, according to CORVUS ISR’s published matrix. In the dense configuration with 400 movers, the count declined from 14,032 to 8,040 per minute.
The benchmark holds the synthetic scene, sensor model, detections and metric definitions constant while changing only the tracker. Each row uses seed 1337, a 20-second warm-up and a 120-second measurement period. Because the demonstration contains no real people, vehicles or locations, its generated ground truth records each simulated object’s identity exactly.
CORVUS ISR also reported smaller gains under harsher conditions: 16.6% fewer switches at 0.5 fps, 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 are identical between models by design, according to the benchmark description.
Identity Stability Improves Under Load
Fewer identity switches can make tracked movement easier to interpret because the software is less likely to assign a new label to the same object or confuse nearby objects. The reported reduction under both baseline and dense traffic suggests the newer association method handles crowded synthetic scenes better than the deliberately simple baseline.
The performance data also indicate that the added tracking logic remained within the demonstration’s processing limit. At 400 simulated movers, v2 averaged about 1.2 milliseconds per sensor tick and reached roughly five milliseconds in its slowest reported result, against a 10-millisecond budget. Those figures apply to the browser demonstration and do not establish performance on operational imagery or other hardware.
As an affiliate, we earn on qualifying purchases.
Fixed Seed Exposes Tracker Differences
The archived v1 tracker, described as greedy nearest-neighbour, uses two-pass greedy association, constant-velocity prediction and fixed two-second coasting. The v2 model, called confirmed-track auction, adds track confirmation, three-tier auction association, velocity-consistency gating, a noise-scaled reservation price and confidence-decayed coasting.
CORVUS ISR uses a stricter identity-switch measure than the standard MOTChallenge definition. Its benchmark counts every change in the track identity assigned to a ground-truth object, including fragmentation and reacquisition. That produces high absolute error totals and means the published numbers should not be compared directly with results using a narrower definition.
“Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”
— CORVUS ISR benchmark publication
Outside Replication Is Still Pending
It is not yet clear whether independent researchers have reproduced the results or whether the same gains would hold on real aerial imagery. The benchmark is published and runnable without registration, but it remains hosted and described by CORVUS ISR.
The tracker still produced thousands of identity errors per minute in several stress tests. The publication also says an AI executor built the tracker under a written acceptance contract and that it was independently reviewed before release, but the reviewer’s identity, review method and findings were not provided.
Future Trackers Face Same Seed
CORVUS ISR says future tracker versions will be added as new rows against the same fixed seed, allowing readers to compare later changes with the current baseline. Users can also select Run benchmark in the public demonstration to rerun the present matrix without an account or nondisclosure agreement.
The next meaningful test will be whether outside parties reproduce the measurements and whether later evaluations use different seeds, real-world datasets or other tracking metrics. Those results would show how far the reported gain extends beyond this controlled synthetic scene.
Source: Thorsten Meyer AI
Key Questions
What did CORVUS ISR report?
The company reported 42.1% fewer identity switches with 150 movers and 42.7% fewer with 400 movers when comparing its v2 tracker with the v1 baseline.
What is an identity switch?
An identity switch occurs when tracking software changes the track label assigned to the same object. CORVUS ISR’s stricter metric also counts fragmentation and reacquisition as switches.
Does the benchmark use real surveillance footage?
No. CORVUS ISR describes the demonstration as entirely synthetic, with no real people, vehicles or places. Generated scenes provide exact ground-truth identities for measurement.
Can the public reproduce the test?
CORVUS ISR says users can run the fixed-seed benchmark in a browser without signup or an NDA. That provides access to the same company-hosted test, but independent replication has not been documented.
Did v2 eliminate tracker identity errors?
No. Despite the reported percentage reductions, v2 still recorded 1,183 switches per minute in the baseline configuration and 8,040 per minute in the 400-mover test.
Source: Thorsten Meyer AI
