- Under stadium floodlights a bald head shines like a ball. In Oct 2020 a Pixellot auto-camera locked onto a linesman's head instead of the ball for a whole match. The detector scored the head 98% confident. The pixels lied. We built the layer that overrules them. 🧵
- A detector scores each frame on texture and shape, with no memory of where a real ball can be next. So a shiny head can outscore the actual ball. And raising the confidence knob still grabs the impostor while dropping real low-confidence balls. It's the wrong knob.
- Plumbline sits downstream of any detector and runs three deterministic physics gates the model cannot override: a Kalman/UKF kinematic gate, an optical-flow consistency gate, and a geometric perspective gate. Plain numpy/scipy code outside any model.
- The kinematic gate turns each candidate's innovation into a Mahalanobis distance. Accept under 3 sigma, reject over 5. On the bald_linesman scenario (synthetic, seed 7) the head impostor scored up to 82.3 sigma. Physically impossible. Rejected.
- Over that 44-frame clip the simulated detector fired 18 high-confidence false balls on the head. The gates rejected all 18. The real ball, at 0.87 confidence and 0.4 sigma, passed all three gates and stayed locked. Physics overrules pixels.
- The bald head is an author-injected reproduction of the 2020 incident, not organic detector noise. We scripted the exact failure that never shows up in a detector's test set, then proved the gate mechanism rejects a physically impossible high-confidence detection.
- Same gate code, two kinds of input. We also ran it on a real broadcast clip with real YOLO11x detections. Honestly image-space (a 2-D Kalman filter), because one uncalibrated clip carries no metric depth. Faking a 3-D filter would have been dishonest.
- On that 240-frame real clip the same gates held the ball 68.3% of frames vs 43.8% for follow-highest-confidence and 30.0% for raised-threshold. Lower than the synthetic 97.7% ceiling by design. Real footage is harder, and we show the honest number.
- These are tracking-quality metrics, detector-quality-independent. A "ball" at a constant 1.7m height moving 3mph is wrong at any detector accuracy. So the gate keeps its value even against a perfect future detector. The failure is physics, not detection.
- As vision goes agentic, what enforces that a detection is physically possible before an agent swings the camera or moves the robot? A confidence score can't. We put that check in plain, inspectable code. #ComputerVision #KalmanFilter #SportsTech #MLOps
- Run Plumbline yourself: the same gates over the synthetic scenes and the real clip, every decision with a physics reason on file. https://veriprajna.com/demos/physics-constrained-computer-vision
Published on X · September 4, 2026
On social media
See this post on its original platform
In our archive