VK-SITS: a Robust Time-Surface for Fast Event-Based Recognition

Event-based cameras are non-conventional sensors that offer movement perception with high dynamic range, high temporal resolution, high power efficiency, and low latency. Nevertheless, because event data are asynchronous and sparse, traditional machine learning and deep learning tools are not suited for this data format. A common practice in event representation learning is to generate image-like representations, usually referred to as time-surfaces. In this paper, we focus on the VK-SITS representation, an end-to-end trainable, spatial and speed-invariant time-surface. We perform additional experiments and an analysis of the influence of meta-parameters to show that VK-SITS is a generic event representation by evaluating it on a new recognition task (SL-Animals-DVS), and give additional intuitions to the choice of its meta-parameters. Results show that VK-SITS is a generic event representation for which optimization of parameters is robust regardless of the split utilized to optimize parametrization.