Abstract
Event cameras, inspired by biological vision systems, provide a natural and data efficient representation of visual information. Visual information is acquired in the form of events that are triggered by local brightness changes. However, because most brightness changes are triggered by relative motion of the camera and the scene, the events recorded at a single sensor location seldom correspond to the same world point. To extract meaningful information from event cameras, it is helpful to register events that were triggered by the same underlying world point. In this work we propose a new model of event data that captures its natural spatio-temporal structure. We start by developing a model for aligned event data. That is, we develop a model for the data as though it has been perfectly registered already. In particular, we model the aligned data as a spatio-temporal Poisson point process. Based on this model, we develop a maximum likelihood approach to registering events that are not yet aligned. That is, we find transformations of the observed events that make them as likely as possible under our model. In particular we extract the camera rotation that leads to the best event alignment. We show new state of the art accuracy for rotational velocity estimation on the DAVIS 240C dataset. In addition, our method is also faster and has lower computational complexity than several competing methods.
Introduction
Unlike traditional frame-based cameras that capture intensity at fixed intervals, event cameras (such as the Dynamic Vision Sensor, DVS) record pixel-level brightness changes asynchronously. Each event is a tuple $(x, y, t, p)$ representing the spatial coordinates, timestamp, and polarity of the change. While this paradigm offers high temporal resolution and low power consumption, it introduces a significant challenge: motion blur. When the camera moves relative to the scene, a single world point triggers a sequence of events across different pixels over time. To recover scene structure or estimate camera motion, these events must be aligned—a process of "de-warping" the spatio-temporal event stream so that events originating from the same world point map to the same spatial coordinate.
Methodology: The ST-PPP Framework
The proposed method treats the event stream as a Spatio-Temporal Poisson Point Process (ST-PPP). This probabilistic approach moves beyond heuristic "contrast maximization" techniques toward a rigorous statistical foundation.
Generative Model and Likelihood Maximization
The core intuition is that in a perfectly aligned state, the number of events at any given pixel should follow a Poisson distribution. Pixels corresponding to high-contrast edges in the environment will exhibit high event rates, while homogeneous regions will have rates near zero. The goal is to find the transformation parameters (e.g., rotational velocity $\hat{\omega}$) that maximize the likelihood of the observed events $\mathcal{O}$ under the model.
Handling Unknown Scene Parameters
A significant challenge in event alignment is that the "true" event rate $\lambda$ for each pixel is unknown. To address this, the authors employ a partially Bayesian approach, modeling $\lambda$ with a Gamma distribution and marginalizing it. This results in a Negative Binomial distribution for the event counts, allowing for optimization without a prior map of the scene. The Adam optimizer is then used to maximize the joint log-likelihood of the positive and negative event streams independently.
Experimental Evaluation
The method was evaluated using the DAVIS 240C Dataset, which includes sequences such as boxes, poster, dynamic, and shapes. Ground truth was provided by high-frequency IMU data and motion capture poses.
Key Results
The ST-PPP model achieved state-of-the-art accuracy in rotational velocity estimation. Notably, it demonstrated a 39% improvement on the shapes sequence. Sparse scenes like shapes are traditionally difficult for contrast-based methods due to the lack of dense texture; however, the probabilistic nature of ST-PPP handles such sparsity effectively.
Furthermore, the method is computationally efficient. With a complexity of $O(N_e)$ relative to the number of events, it is significantly faster than entropy-based methods that scale quadratically. A single loss function evaluation takes approximately 3.1ms on a standard CPU.
Conclusion
The ST-PPP framework provides a simple yet powerful probabilistic model for event camera data. By shifting from heuristics to likelihood maximization, the authors achieve superior accuracy and efficiency. Future work includes extending this model from 3-DOF rotation to full 6-DOF motion, accounting for translation and depth-induced scale changes in the event stream.
Summary updated