In the realm of wide-area motion imagery (WAMI), maintaining accurate object identity over time is a critical challenge. The latest benchmark from Corvus ISR compares two tracker models in a synthetic scene with perfect ground truth, providing a clear view of their performance. The key metric, ID switches per minute, exposes how often the tracker confuses object identities, a crucial factor in surveillance reliability.
The baseline model, v1 “greedy nearest-neighbour,” employs a straightforward two-pass greedy association with constant velocity prediction and fixed 2-second coasting. Despite its simplicity, it still manages to process scenes with up to 400 movers at 2 frames per second, serving as a solid foundation for comparison. In contrast, v2 introduces an auction-based confirmation approach, utilizing a three-tier auction association, velocity-consistency gating, and noise-scaled reservation pricing to refine track accuracy.
When tested under typical conditions, the v1 model registered 2,042 ID switches per minute in a scenario with 150 movers. The v2 model demonstrated a significant reduction, bringing this number down to 1,183 — a remarkable 42.1% improvement. Similar gains appeared in denser scenes with 400 movers, dropping from 14,032 to 8,040 switches (−42.7%). These results underscore how advanced association strategies can dramatically reduce identity errors.
Even under challenging conditions such as frame starvation (0.5 fps), occlusion (20%), or degraded sensor quality, the v2 tracker consistently cut the number of switches by around 18%. Importantly, these metrics are measured against perfect ground truth, making them a true reflection of model robustness. It’s worth noting that the detection rate remains identical for both models, highlighting that improvements stem purely from the tracking algorithms themselves.
From an engineering perspective, v2 operates at roughly 1.2 milliseconds per sensor tick at 400 density, well within real-time constraints. The entire process is accessible through the live demo, where anyone can reproduce the benchmark in their browser—no registration or NDA required. This transparency is intentional; by openly publishing these failure metrics, Corvus ISR emphasizes that every new tracker must prove itself against the same standard.
In essence, this synthetic benchmark showcases how a sophisticated, auction-based multi-object tracker can reduce identity errors by over 40%, all while running seamlessly in real time. For tech enthusiasts and developers alike, the ability to see the public benchmark and reproduce it live provides invaluable insight into the power of modern AI-driven tracking systems. Feel free to run the benchmark yourself and witness the results firsthand.
As an affiliate, we earn on qualifying purchases.
wide-area surveillance camera system
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
real-time motion detection camera
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
