Cross-Domain RGB-to-Event Image Translation Using Spiking Neurons for Eye Detection

초록

Event cameras possess a unique ability to accurately detect rapidly moving objects, making them valuable in various fields. However, in eye detection using event cameras, there is a scarcity of public datasets with annotations, and the low spatial information characteristic of event images makes manual labeling challenging. To address this, we propose an RGB-to-event translation method using spiking neurons from a single RGB image. Our approach starts by applying the adaptive StyleFlow algorithm to generate event-style images. It is then followed by a Leaky Integrate-and-Fire (LIF) neuron that captures the asynchronous pixel firing behavior of event cameras, and a region-aware polarity generation module that produces realistic event images. The eye detection process utilizes the RetinaFace algorithm through cross-modal learning by combining the generated synthetic event images with existing RGB datasets. To demonstrate the effectiveness of our method, we evaluated both a custom self-collected dataset and the public N-Caltech101 dataset. Results showed an accuracy of 96.52% in face detection and 98.26% in eye detection on our self-collected dataset. On the public N-Caltech101 dataset, the method achieved an accuracy of 96.62% in face detection and 98.31% in eye detection. This advancement demonstrates the robustness and effectiveness of our method, outperforming the existing state-of-the-art event camera-based eye detection methods.

키워드