In the realm of **wide-area motion imagery (WAMI)**, maintaining accurate object identities across frames is a persistent challenge. Corvus ISR has published a compelling public tracker benchmark that compares two models on a synthetic scene with perfect ground truth, offering an insightful look into how advanced tracking algorithms perform under stress. This benchmark uses identical detection and sensor parameters, isolating the tracker as the sole variable in the performance results.
The first model, **v1**, employs a straightforward *greedy nearest-neighbour* approach, featuring a two-pass association with fixed velocity prediction and a fixed 2-second coasting period. Despite its simplicity, it serves as the published baseline. The newer **v2** model introduces a sophisticated *auction-based confirmation* system, including three-tier association, velocity gating, and confidence decay. This design aims to reduce identity errors, which are critical in surveillance and tracking tasks where object identities must be reliably maintained over time.
The results are striking: in a scenario with 150 movers at 2fps, the number of ID switches per minute dropped from 2,042 to 1,183—an impressive 42.1% reduction. When scaled to a dense scene with 400 objects, switches decreased from 14,032 to 8,040, a 42.7% improvement. These metrics, measured under strict conditions—including occlusion, noise, and jitter—highlight the robustness of the auction-based approach in complex environments.
One notable aspect of this benchmark is its transparency: the published numbers are raw, unfiltered error counts. The ID switch metric used here is intentionally strict, counting every change in the assigned identity—even re-acquisitions and fragmentations—thus providing a rigorous measure of tracker fidelity. Despite these improvements, both models still generate thousands of identity errors per minute, confirming that multi-object tracking remains a challenging problem.
From an engineering perspective, the v2 model operates at approximately 1.2 milliseconds per sensor tick at a density of 400 objects, comfortably within real-time constraints. This performance is achieved in a browser environment, with no sign-up or NDA required—anyone can reproduce it live and see the results firsthand. The entire process was orchestrated by an AI executor, reviewed independently, and tested against synthetic data with pixel-perfect ground truth.
Corvus ISR emphasizes that all scenes are synthetic—no real persons, vehicles, or locations are involved—making these measurements a pure test of algorithmic performance rather than environmental factors. The company is committed to transparency, publishing both successes and failures to foster genuine progress in multi-object tracking technology. If you’re curious to see how these algorithms perform firsthand, run the benchmark yourself and explore the capabilities of this innovative system.
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.
