CORVUS ISR is a wide-area motion imagery (WAMI) exploitation product built entirely as a synthetic demonstration. Every pixel is generated: no real persons, vehicles, or places appear anywhere in the product.
Multi-object tracking is about preserving an object’s assigned identity across successive frames, making identity switches a critical measure of tracker performance. CORVUS ISR published the public benchmark using an identical fixed-seed synthetic scene with perfect ground truth: seed 1337, a 20s warm-up, and 120s measured per row. The sensor model, detection generation, and metric definitions are byte-identical; only the tracker differs.
The v1 model, “greedy nearest-neighbour,” is the deliberately simple baseline and published floor: two-pass greedy association, constant-velocity prediction, and fixed 2s coasting. It remains runnable in archived demo slices 1-2. The current v2 model, “confirmed-track auction,” appears in demo slice 3 and adds track confirmation, three-tier auction association, velocity-consistency gating, a noise-scaled reservation price, and confidence-decayed coasting.
In the baseline configuration of 150 movers at 2fps, ID switches per minute fell from 2,042 to 1,183, a 42.1% reduction. In the dense configuration with 400 movers, they fell from 14,032 to 8,040, a 42.7% reduction.
The gains remained measurable under other stresses: 16.6% fewer switches in the frame-starved 0.5fps row, 18.6% fewer with 20% occlusion, and 18.1% fewer in the degraded 1fps + jitter, 70% contrast row. Detection rate is a sensor property and is identical for both models by construction.
The benchmark uses a deliberately stricter metric than the MOT-challenge IDSW definition, counting every change of the track identity assigned to a ground-truth object. Fragmentations and re-acquisitions count as switches too. Both models still commit thousands of identity errors per minute under stress, and CORVUS ISR publishes those rows on its own site because synthetic scenes provide perfect ground truth, making these measurements rather than marketing.
At a density of 400, v2 averages ~1.2ms per sensor tick, with a worst result of ~5ms against a 10ms budget, delivering real-time performance in the browser. The tracker was built by an AI executor against a written acceptance contract and independently reviewed before shipping.
The publication principle is explicit: “Vendors who show only successes ask for faith; a published failure matrix asks for measurement.” Every future tracker must land as a new public row against the same seed, and every current row can be run without signup or an NDA. Open the demo, press “Run benchmark,” and reproduce it live to run the benchmark yourself.
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.
ZOSI 3K Lite Security Camera System Indoor Outdoor,AI Human/Vehicle Detection,Night Vision,Remote Access,4pcs 1080P 1920TVL Weatherproof Surveillance Cameras,H.265+ 8CH CCTV DVR(NO HDD)
- High-Quality 8CH DVR: Supports 3K Lite and 1080P cameras
- Weatherproof Surveillance Cameras: Indoor/outdoor use with rust-resistant ABS
- Infrared Night Vision: Up to 80 feet in darkness
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Bouncie GPS Tracker for Vehicles with Real-Time Location, Route History, Speed Monitoring, Geo-Fence & Accident Notifications – for Fleets & Family – Cancel Anytime
- Real-Time Location Updates: Up-to-the-minute vehicle location tracking
- Unlimited Trip History: Detailed route and trip records
- Speed and Driving Behavior Monitoring: Tracks speed, acceleration, and braking
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
