An Interpretable Pixel Intensity Reconstruction Model for Asynchronous Event Camera

Event cameras with high temporal resolution and high dynamic range have great potential in computer vision (CV) tasks. To utilize the deep neural networks directly, an efficient reconstruction method converting event-based data to frame-based is necessary. In this work, the interpretable Event Represented Intensity (ERI) model that recovers the logarithm of the intensity sensed by a dynamic vision pixel is proposed for the first time. The amplitude-frequency characteristic of the recovered logarithm of the intensity is used to construct the frame-based image to complete CV tasks. Experiment results on the N-Caltech101 dataset show that the proposed ERI model achieves the classification accuracy of 79.20%, which balances the performance and computation cost better.