Image to Events: An Event Simulator Combining Circuit Characteristics and Domain Adaptation

Due to the novelty of the sensor, event camera applications often rely on simulated event data for training. However, most existing simulators overlook the inherent circuit characteristics of event cameras and the potential noise introduced by such hardware properties, while a domain shift also persists between simulated and real event data. To address these issues, we propose an image-to-event simulator (ITES). First, we model the voltage signal inside an event camera as a Brownian motion with drift and introduce Brownian stochastic terms correlated with temperature and brightness to simulate noise. Second, we employ unsupervised domain adaptation (UDA) to reduce the domain shift between simulated and real events. Finally, we also propose a novel event simulation evaluation metric (ESEM), which enables quantitative assessment of simulation results. Experimental results demonstrate that, compared to those from other simulation methods, the events simulated by our method achieve superior performance in both visual quality and in quantitative metrics, and exhibit a closer resemblance to real events. Furthermore, in tests of downstream applications using event cameras, networks trained on our simulated events achieve better generalization on real event data.