Noise2Params: Calibrate event cameras using only static noise

Event cameras (ECs) are neuromorphic sensors that output streams of per-pixel intensity changes, enabling high-speed, low-power vision. However, their calibration has been tricky—requiring specialized dynamic light sources and complex procedures. Now, a new paper from Carnegie Mellon researchers (Root et al.) introduces Noise2Params, a foundational probabilistic model that for the first time unifies static-scene noise events and step response curves (S-curves) within a single analytical framework grounded in photon statistics. The model yields three formulations covering all intensity regimes: exact Poisson, saddle-point, and Gaussian.

From this model, the team derives a practical method that extracts three essential camera parameters—log-contrast threshold B, lux-to-photon conversion factor α, and intensity-dependent leakage term θ—by minimizing error against observed noise-event distributions. The key advantage: Noise2Params requires only simple recordings of static, uniform scenes, making calibration experimentally accessible to anyone with an event camera. To validate, they trained convolutional neural networks on synthetic noise images generated from their distributions and found that models incorporating synthetic data significantly outperformed those trained solely on real experimental data. This opens the door to robust, noise-aware algorithm design and calibration in photon-limited regimes, with implications for autonomous driving, robotics, and high-speed imaging.