EventRPG: Event Data Augmentation with Relevance Propagation Guidance

• DOI: 10.48550/arxiv.2403.09274
• Publication Date: 2024-03-14
• Authors: Mingyuan Sun; Donghao Zhang; Zongyuan Ge; Jiaxu Wang; Jia Li; Zheng Fang; Renjing Xu
• Keywords: Relevance; Event data

Event camera, a novel bio-inspired vision sensor, has drawn lot of attention for its low latency, power consumption, and high dynamic range. Currently, overfitting remains critical problem in event-based classification tasks Spiking Neural Network (SNN) due to relatively weak spatial representation capability. Data augmentation is simple but efficient method alleviate improve the generalization ability neural networks, saliency-based methods are proven be effective image processing field. However, there no approach available extracting saliency maps from SNNs. Therefore, first time, we present Layer-Time-wise Relevance Propagation rule (SLTRP) Layer-wise (SLRP) order SNN generate stable accurate CAMs maps. Based on this, propose EventRPG, which leverages relevance propagation spiking network more augmentation. Our proposed been evaluated several structures, achieving state-of-the-art performance object recognition including N-Caltech101, CIFAR10-DVS, with accuracies 85.62% 85.55%, as well action task SL-Animals an accuracy 91.59%. code at https://github.com/myuansun/EventRPG.