Abstract
Event cameras offer many advantages over standard cameras due to theirdistinctive principle of operation: low power, low latency, high temporalresolution and high dynamic range. Nonetheless, the success of many downstreamvisual applications also hinges on an efficient and effective scenerepresentation, where Neural Radiance Field (NeRF) is seen as the leadingcandidate. Such promise and potential of event cameras and NeRF inspired recentworks to investigate on the reconstruction of NeRF from moving event cameras.However, these works are mainly limited in terms of the dependence on dense andlow-noise event streams, as well as generalization to arbitrary contrastthreshold values and camera speed profiles. In this work, we propose Robuste-NeRF, a novel method to directly and robustly reconstruct NeRFs from movingevent cameras under various real-world conditions, especially from sparse andnoisy events generated under non-uniform motion. It consists of two keycomponents: a realistic event generation model that accounts for variousintrinsic parameters (e.g. time-independent, asymmetric threshold andrefractory period) and non-idealities (e.g. pixel-to-pixel thresholdvariation), as well as a complementary pair of normalized reconstruction lossesthat can effectively generalize to arbitrary speed profiles and intrinsicparameter values without such prior knowledge. Experiments on real and novelrealistically simulated sequences verify our effectiveness. Our code, syntheticdataset and improved event simulator are public.