EvDownsampling: A Robust Method for Downsampling Event Camera Data

Abstract

Event cameras are bio-inspired vision sensors that are well-suited for use on autonomous systems. Compared to conventional cameras, event cameras provide a higher dynamic range and temporal resolution as well as lower bandwidth and power requirements. Furthermore, when the events from such cameras are processed by spiking neural networks (SNNs) running on neuromorphic chips, very energy-efficient systems can be implemented. However, the spatial resolutions offered by modern event-based cameras is much higher than most neuromorphic hardware can process and thus, it is often necessary to downsample the event stream. While downsampling is regularly used in standard computer vision, event-based downsampling algorithms are lacking. This hampers the development of event-based computer vision systems. Here, we present a bio-inspired spatio-temporal downsampling technique that can downsample event streams by factors of up to 16 times. To evaluate our downsampling method, we compare downsampled event streams from a high-resolution event camera to those produced natively by a low-resolution event camera. Finally, we show that compared to other event-based downsampling techniques, we can achieve improved accuracy with considerably lower numbers of spikes when classifying highly downsampled versions of the DVSGesture dataset. The code is available at: https://github.com/anindyaghosh/EvDownsampling and the event streams can be downloaded from: https://doi.org/10.25377/sussex.26528146 .