Data Augmentation in Temporal and Polar Domains for Event-Based Learning

Event cameras are inherently suitable for spikingneural networks (SNNS) and have great potential in challenging scenesdue to the advantages of bionics, asynchrony, high dynamic range, and no motion blur.However, novel data augmentations designed for event properties are required to process the unconventional output of these cameras in order to unlock their potential.In this paper, we explore the extraordinary influence of brightness variations due to event properties. Along the way, two novel data augmentation methods, lemph[EventInvert) and lemph(EventDrift) (EventID), are proposedto simulate two basic transformations of this influence.Specifically, EventID inverts or drifts events in the stream through transformationsin temporal and polar domains, thereby generating samples affected by brightness variances.Extensive experiments are carried out on the CIFAR10-DVS, N-Caltech101, and N-CARS datasets.It turns out that this simulation improves generalization by increasing the robustness of models against brightness variations.In addition, EventID is broadly effective, surpassing previous state-of-the-art performances.For example, the spiking neural network model with EventID achieves a state-of-the-art accuracy of 83.501% on the CIFAR10-DVS dataset.