TL;DR
Corvus ISR has published a reproducible synthetic benchmark showing that its v2 multi-object tracker recorded about 42% fewer identity switches than its v1 baseline in standard and dense tests. The figures are company-published results rather than external validation, and both trackers still produced thousands of errors per minute under stress.
Corvus ISR has published a reproducible synthetic benchmark reporting that its v2 multi-object tracker cut identity switches by 42.1% in a 150-mover test and 42.7% in a 400-mover test compared with its deliberately simple v1 baseline. The results matter because identity switches measure whether tracking software can preserve the correct assigned identity as objects move through successive frames.
In the baseline configuration, which used 150 moving objects at two frames per second, reported identity switches fell from 2,042 to 1,183 per minute. In the denser 400-mover configuration, the count declined from 14,032 to 8,040 per minute. Corvus ISR says each row used the same fixed-seed scene, sensor model, detections and metric definitions, with only the tracker changing.
The test scene used seed 1337, followed by a 20-second warm-up and 120 seconds of measurement for each row. Corvus ISR is a wide-area motion imagery exploitation demonstration made entirely with synthetic imagery; the company says no real people, vehicles or locations appear in it. Synthetic generation provides exact ground truth against which assigned track identities can be checked.
The reported improvement was smaller under other test conditions. Corvus ISR 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 equal by construction because detection generation was not changed between tracker runs.
Fewer Errors in Dense Tracking
An identity switch occurs when a ground-truth object becomes associated with a different track identity. Reducing these events can improve trajectory continuity, limit confusion between nearby objects and make downstream analysis more dependable. The larger reductions in the 150- and 400-mover configurations indicate that v2 handled association decisions better than the baseline under those specified test conditions.
The benchmark also reports browser-level processing speed. At a density of 400 objects, v2 averaged about 1.2 milliseconds per sensor tick, with a reported worst result of about five milliseconds against a 10-millisecond processing budget. If reproduced on comparable hardware, that leaves the tracker within the demonstration’s real-time target while applying more complex association logic.
multi-object tracking software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Two Trackers Differ
The archived v1 tracker, called “greedy nearest-neighbour”, uses two-pass greedy association, constant-velocity prediction and fixed two-second coasting. It was designed as a simple published baseline and remains available in demo slices one and two.
The current v2 tracker, called “confirmed-track auction”, is used in demo slice three. It adds track confirmation and three-tier auction association, along with velocity-consistency gating, a noise-scaled reservation price and confidence-decayed coasting. Thorsten Meyer AI says an AI executor built the tracker from a written acceptance contract and that it received independent review before release; no details about the reviewer or review procedure were supplied.
Corvus ISR applies a stricter identity-switch measure than the MOTChallenge IDSW definition. Its metric counts every change in the identity assigned to a ground-truth object, including fragmentation and reacquisition events. That choice produces higher absolute totals and means direct comparisons with benchmarks using other definitions may be misleading.
“Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”
— Corvus ISR publication principle
AI object tracker for surveillance
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
External Validation Is Still Missing
The benchmark is published by Corvus ISR, and no results from an unaffiliated laboratory or public replication are included in the available information. It is also unclear how the tracker would perform on real sensor imagery, where calibration errors, environmental variation and imperfect ground truth may differ from a controlled synthetic scene.
Both versions continued to produce thousands of identity errors per minute in the reported stress tests. The percentage reductions describe improvement over a deliberately basic baseline, not error-free tracking. The available results also do not establish whether v2 would outperform other modern tracking methods, since the matrix compares only the two Corvus ISR versions.
synthetic imagery analysis tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Future Versions Face the Same Seed
Corvus ISR says every future tracker version will be added as a new public benchmark row using the same seed and test construction. Readers can visit the benchmark and demonstration pages, select “Run benchmark” and compare the displayed results with the published matrix. Independent replications, tests on other scenes and comparisons with outside trackers would provide stronger evidence about how broadly the gains apply.
real-time object tracking hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is an identity switch?
An identity switch occurs when the tracking system changes the track identity assigned to a ground-truth object. Corvus ISR also counts fragmentation and reacquisition as switches.
How large was the reported improvement?
Corvus ISR reported 42.1% fewer switches with 150 movers and 42.7% fewer with 400 movers. Reductions under lower frame rates, occlusion and degraded imagery ranged from 16.6% to 18.6%.
Does the benchmark use real surveillance imagery?
No. The product is an entirely synthetic demonstration. Corvus ISR says every pixel is generated and that no real people, vehicles or places are shown.
Can the results be independently reproduced?
The test can be rerun through the public browser demonstration without registration or an NDA. That supports direct checking of the published rows, but independent external validation has not been provided in the available material.
Does v2 eliminate tracking errors?
No. Although v2 produced fewer identity switches than v1, it still recorded thousands of switches per minute under demanding configurations. The results show a relative reduction against the project’s baseline, not complete identity preservation.
Source: Thorsten Meyer AI
Source: Thorsten Meyer AI
