Imagine you're trying to tune four guitars at once in a noisy bar. Each string vibrates at its own frequency, but all four guitars share the same room acoustics — the reverb, the standing waves, the background chatter. If you could model the room once and subtract it from all four signals simultaneously, each individual tuning problem gets dramatically easier. That is the core mechanism of CoHelix: a single shared geometric model (a homography describing how the camera moves relative to the propeller plane) informed by blade events from all four rotors, making each individual rotation estimate more robust. The committed claim: by coupling four coplanar propeller trackers through one shared projective geometry, CoHelix achieves 62.9% lock on moving-camera data — meaning it holds within 2% of ground truth for nearly two-thirds of evaluated time — compared to 39.8% for HelixTrack and 22.2% for EventPro. On a separate target-follower dataset with a different quadcopter, lock reaches 95.9% versus 56.2% and 61.1% for the same competitors. This is not a marginal improvement; it is a structural redesign of the estimation pipeline. The architecture is a recursive filter — think extended Kalman filter territory — that maintains one homography state shared across all propellers plus four independent rotation states. Each incoming event from the event camera is assigned to a propeller, updates that propeller's rotation estimate, and its geometric residual corrects the shared homography. The key insight is that propellers on a rigid quadcopter frame are coplanar, so a single projective transformation describes camera-to-plane geometry for all four. This is elegant because it converts what would be four independent, geometry-starved tracking problems into one well-observed geometry problem with four frequency-tracking subproblems. The paper introduces "lock" as a primary metric — the fraction of time the estimate stays within 2% of synchronized ground truth — and argues convincingly that mean error alone hides catastrophic failures. An estimator with low mean error could be wildly wrong half the time and perfect the other half. Lock tells you how reliable the tracker actually is. The ablation is devastating: with only one geometry contributor instead of four, lock drops from 62.9% to 3.4%. The shared geometry is not a nice-to-have; it is the load-bearing wall. Validation uses leave-one-recording-out evaluation on moving-camera data, plus a separate private dataset with a different target drone. The cross-dataset transfer — using measurement settings tuned on one dataset applied directly to the other — is a genuine robustness test. The baselines (HelixTrack, EventPro) are the named prior art in this niche. That said, the private dataset limits independent replication, and the evaluation is entirely the authors' own. Where this matters: autonomous drone-tracking systems need to estimate target propeller rates for identification, intent estimation, and threat classification. Event cameras are the natural sensor for high-speed rotation because they respond to brightness changes at microsecond resolution rather than fixed frame rates. The bottleneck has been disentangling blade events from camera ego-motion — exactly what CoHelix's shared homography addresses. The path forward is real-time onboard deployment and scaling to scenarios with multiple target drones, each contributing its own propeller-plane geometry. The paper is a clean engineering contribution to a narrow but important problem. It does not claim generality beyond coplanar multi-propeller tracking, and the mechanism — shared projective constraints reducing per-tracker ambiguity — is well-motivated and well-ablated. The 18.5× lock improvement from the shared geometry ablation alone justifies the architectural choice.
