A Spatio-Temporal Event Data Augmentation Method for Dynamic Vision Sensor

Abstract

The advantages of Dynamic Vision Sensor (DVS) camera and Spiking Neuron Networks (SNNs) have attracted much attention in the field of computer vision. However, just as many deep learning models, SNNs also have the problem of overfitting. Especially on DVS datasets, this situation is more severe because DVS datasets are usually smaller than traditional datasets. This paper proposes a data augmentation method for event camera, which augments asynchronous events through random translation and time scaling. The proposed method can effectively improve the diversity of event datasets and thus enhance the generalization ability of models. We use a Liquid State Machine (LSM) model to evaluate our method on two DVS datasets recorded in real scenes, namely DVS128 Gesture Dataset and SL-Animals-DVS Dataset. The experimental results show that our proposed method improves baseline accuracy without any augmentation by 3.99% and 7.3%, respectively.