The Impact Of Corvus ISR On AI Tracker Performance: 42% Fewer ID Switches

CORVUS ISR has published a reproducible synthetic benchmark reporting that its current v2 multi-object tracker made 42.1% fewer identity switches than its v1 baseline with 150 moving objects and 42.7% fewer with 400. The result matters because identity continuity is a core test of whether an AI tracker can follow the same object across successive frames, but the figures remain self-published benchmark results rather than an independent evaluation.

In the benchmark’s baseline configuration of 150 movers at two frames per second, identity switches fell from 2,042 to 1,183 per minute. In the denser 400-mover configuration, the reported rate declined from 14,032 to 8,040 per minute. CORVUS ISR says every row uses the same fixed-seed synthetic scene, seed 1337, with a 20-second warm-up and 120-second measurement period.

The company describes v1, called “greedy nearest-neighbour,” as a deliberately simple baseline using 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.

Smaller gains were reported under other stresses: 16.6% fewer switches at 0.5 fps, 18.6% fewer with 20% occlusion and 18.1% fewer in a degraded one-frame-per-second test with jitter and 70% contrast. Detection rates are identical by design because the sensor model and detection generation do not change between tracker runs.

Identity Continuity Improves, Errors Persist

An identity switch occurs when a tracker assigns a different track identity to the same ground-truth object. Reducing these errors can improve trajectory analysis, movement histories and other systems that depend on keeping identities stable through crowding, missed detections or occlusion.

The reported reductions show that association logic alone can change tracking performance when detections remain fixed. Yet the absolute totals temper the result: v2 still produced 1,183 switches per minute in the baseline test and 8,040 per minute with 400 movers. The benchmark supports a comparative claim about these two implementations, not a claim that identity tracking has been solved.

A Fixed Synthetic Test Bed

CORVUS ISR is presented as a wide-area motion imagery exploitation demonstration built entirely from synthetic data. Its documentation says no real people, vehicles or locations appear in the product, allowing every simulated object’s identity to be known throughout a run and providing perfect ground truth for scoring.

The benchmark uses a stricter definition than the common MOT Challenge identity-switch measure. It counts every change in the identity assigned to a ground-truth object, including fragmentations and reacquisitions. Archived demo slices retain the v1 tracker, while demo slice 3 contains v2. Thorsten Meyer AI also says the newer tracker was built by an AI executor against a written acceptance contract and reviewed before release.

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

— CORVUS ISR benchmark documentation

Independent Replication Is Still Missing

The material provided does not identify an independent laboratory or outside research group that has reproduced the results. It is also unclear how the tracker would perform on real sensor footage, where calibration errors, unusual motion, weather, compression and detector bias may differ from the synthetic model.

The benchmark reports v2 averaging about 1.2 milliseconds per sensor tick at 400-object density, with a worst result near five milliseconds against a 10-millisecond budget. Those figures indicate browser-based real-time operation in the published setup, but hardware and browser details are not specified in the supplied documentation. No statistical uncertainty across multiple seeds is reported because the comparison uses one fixed seed.

Future Trackers Face the Same Seed

CORVUS ISR says each future tracker will be added as a new public row using the same seed and measurement rules. Readers can also open the public demo and select “Run benchmark” without registration or a nondisclosure agreement.

The next test of the claim will be whether outside users reproduce the published rows and whether later evaluations add more seeds, documented hardware and real-world datasets. Those additions would show how well the reported reduction carries beyond this controlled synthetic comparison.

Key Questions

What does 42% fewer ID switches mean?

It means the v2 tracker changed the assigned identity of a ground-truth object less often than v1 in two published tests. The reduction was 42.1% with 150 movers and 42.7% with 400 movers.

Did the detection system improve between v1 and v2?

No. The benchmark keeps the sensor model and generated detections identical. According to CORVUS ISR, only the tracking method changes, isolating the effect of the association logic.

Does the benchmark use real surveillance footage?

No. CORVUS ISR says the scenes are entirely synthetic and contain no real people, vehicles or places. This provides exact ground-truth identities, though it does not establish performance on real footage.

Can readers reproduce the result?

CORVUS ISR says the benchmark can be run through its public demo using the fixed seed and published configuration, without signup or an NDA. Independent reproduction has not been documented in the supplied information.

Source: Thorsten Meyer AI

Source: Thorsten Meyer AI