TL;DR
Listen free for 30 days with Audible
Thousands of audiobooks and originals — cancel anytime.
Start your free trialAs an affiliate, we earn on qualifying purchases.
Corvus ISR has published reproducible synthetic benchmark results showing its v2 tracker made about 42% fewer identity switches than its v1 baseline in two major configurations. The tracker also improved under lower frame rates, occlusion and degraded imagery, but still recorded thousands of errors per minute in demanding tests.
Corvus ISR has published benchmark results showing that its current v2 multi-object tracker reduced identity switches by about 42% in baseline and dense synthetic tests compared with its v1 model. The public test uses a fixed-seed scene with perfect ground truth, allowing users to rerun the reported comparison in a browser without registration.
In a configuration with 150 moving objects at two frames per second, reported identity switches fell from 2,042 to 1,183 per minute, a 42.1% reduction. With 400 movers, the rate declined from 14,032 to 8,040, or 42.7%, according to the published matrix.
Smaller gains appeared under other stresses. Corvus ISR reported 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 generation was held constant between trackers, meaning the comparison was intended to isolate changes in tracking rather than sensor performance.
The v1 “greedy nearest-neighbour” model uses two-pass association, constant-velocity prediction and fixed two-second coasting. The v2 “confirmed-track auction” model adds track confirmation, tiered auction association, velocity-consistency gating, a noise-scaled reservation price and confidence-decayed coasting.
Identity Stability Improves Under Density
Multi-object tracking depends on preserving the same assigned identity as an object moves between frames. Fewer switches can support more reliable movement histories and reduce the risk that software incorrectly joins separate trajectories. The largest reported reductions occurred in the standard and 400-object configurations, where close proximity makes association harder.
The benchmark also reports that v2 remained within its processing allowance. At 400-object density, Corvus ISR measured about 1.2 milliseconds per sensor tick on average and roughly five milliseconds in the worst run, against a 10-millisecond browser budget. Those figures are publisher-reported and have not been independently replicated here.
Data Association for Multi-Object Visual Tracking (Synthesis Lectures on Computer Vision)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Fixed-Seed Test Design
Corvus ISR is presented as a synthetic wide-area motion imagery demonstration. Its images contain no real people, vehicles or locations. Each benchmark row uses seed 1337, a 20-second warm-up and a 120-second measurement period. Corvus ISR says the sensor model, detections and metric definitions remain byte-identical while only the tracker changes.
The test applies a stricter identity-switch definition than the common MOT Challenge measure. Every change in the track identity assigned to a ground-truth object is counted, including fragmentations and reacquisitions. That produces high absolute error totals but provides a consistent basis for comparing the two models within this test environment.
“Vendors who show only successes ask for faith; a published failure matrix asks for measurement.”
— Thorsten Meyer AI, describing the benchmark’s publication principle
Independent Validation Still Missing
It is not yet clear how closely the gains would carry over to real aerial imagery, different traffic patterns or detector errors not represented by the synthetic generator. The results come from one fixed-seed environment published by the product’s developer, and no outside laboratory assessment was cited.
Both versions also continue to produce thousands of switches per minute in dense or stressful conditions. The relative reductions show an improvement inside the benchmark, but they do not establish that v2 meets the accuracy requirements of any operational deployment. Corvus ISR says an AI executor built the tracker under a written acceptance contract and that it received independent review before release, although details of that review were not provided.
Future Trackers Face Same Seed
Corvus ISR says each future tracker will be added as a new public benchmark row using the same seed and definitions. Readers can test the present claims by opening the demonstration and selecting “Run benchmark”. Broader testing across additional seeds, real-world footage and outside implementations would provide stronger evidence about how well the reported gains generalize.
Key Questions
What improvement did the v2 tracker record?
Corvus ISR reported 42.1% fewer identity switches with 150 movers and 42.7% fewer with 400 movers compared with v1.
What is an identity switch?
An identity switch occurs when tracking software assigns a different track identity to the same ground-truth object. The Corvus ISR test also counts fragmentation and reacquisition events.
Does the benchmark use real surveillance footage?
No. The demonstration is entirely synthetic, with generated pixels, objects and locations. This provides exact ground-truth identities for measurement but leaves real-world performance unproven.
Can users reproduce the reported numbers?
Corvus ISR says users can run the fixed-seed browser benchmark without an account or nondisclosure agreement. Reproducing it confirms behavior in the published test, not performance across independent real-world datasets.
Source: Thorsten Meyer AI
Source: Thorsten Meyer AI
Grilling season Picks
grills
As an affiliate, we earn on qualifying purchases.
