建立统一的事件相机噪声模型，可自动标定关键参数。
Noise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model
- 基于光子统计构建概率模型，统一描述静态噪声与S曲线响应。
- 提出Noise2Params方法，仅需静态均匀场景即可标定3个关键参数。
- 适合事件相机标定、低光环境算法设计者使用。
事件相机（ECs）的精确统一建模仍具挑战，制约校准与算法设计。本文基于光子统计，提出基础的概率模型，统一描述静态场景噪声事件与阶跃响应曲线（S曲线），推导出涵盖全亮度区间的三种概率分布：精确泊松、鞍点近似和高斯分布。模型揭示了看似无关的事件行为背后的内在联系，澄清了S曲线的解读方式，表明其并非简单的固定阈值选择。基于该模型，提出Noise2Params方法，通过最小化误差实现对相机特有参数——对数对比度阈值 $B$、照度到光子转换因子 $α$、泄漏项 $θ$（发现其依赖于光照强度）——的自动标定。该方法仅需静态均匀场景数据，无需特殊动态光源，实验可操作性强。进一步通过合成噪声图像训练卷积神经网络（CNN），并验证其在真实数据上重建静态场景的能力。结果表明，融合合成数据训练的CNN性能优于仅用实测数据训练的模型。本框架为事件相机校准、噪声感知算法设计及弱光条件应用提供了定量基础。
原文摘要 · Abstract (English)
Accurate, unified models for event cameras (ECs) remain elusive, hampering calibration and algorithm design. We develop a foundational probabilistic model for EC event detection, grounded in photon statistics, that unifies the description of static scene noise events and step response curves (S-curves) within a single analytical framework. Three formulations of the probability distributions are derived, spanning all intensity regimes: exact Poisson, saddle-point, and Gaussian. The model reveals the underlying connection between these otherwise disparate EC behaviors and clarifies the interpretation of S-curves, which we show is more nuanced than selecting a fixed probability threshold. Based on this model, we propose Noise2Params, a method for determining camera-specific values of the log-contrast threshold $B$, the lux-to-photon conversion factor $α$, and the leakage term $θ$ (found to be intensity dependent), via error minimization against observed noise-event distributions. Noise2Params requires only recordings of static, uniform scenes, offering an experimentally accessible alternative to approaches that demand specialized dynamic light sources. We further support the validity the model by training convolutional neural networks (CNNs) on synthetic noise images generated from our distributions and evaluating their ability to reconstruct static scenes from experimental data. We further demonstrate the utility of our model by showing that CNNs incorporating synthetic data outperform those trained solely on experimental data. Our framework provides a quantitative foundation for EC calibration, noise-aware algorithm design, and applications in photon-limited regimes.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。