GS2E: Gaussian Splatting is an Effective Data Generator for Event Stream Generation

TL;DR

GS2E tackles the scarcity of large-scale, multi-view, high-fidelity event datasets by reconstructing photorealistic 3D scenes from sparse RGB inputs using 3D Gaussian Splatting and by simulating physically-informed, multi-view event streams. The pipeline combines adaptive, velocity-controlled trajectory generation with a DVS-Voltmeter-based event model that accounts for contrast threshold variability, yielding temporally dense and sensor-faithful events. Key contributions include the end-to-end GS2E pipeline, the dataset of over 1150 scenes with aligned RGB and event streams, and comprehensive validation across 3D reconstruction and image/video tasks, demonstrating strong generalization and realism. This dataset enables robust evaluation and training of event-based methods in realistic, controllable, multi-view settings, reducing reliance on costly real-world captures while highlighting areas for future work such as exposure-aware rendering and dynamic scene modeling.

Abstract

We introduce GS2E (Gaussian Splatting to Event), a large-scale synthetic event dataset for high-fidelity event vision tasks, captured from real-world sparse multi-view RGB images. Existing event datasets are often synthesized from dense RGB videos, which typically lack viewpoint diversity and geometric consistency, or depend on expensive, difficult-to-scale hardware setups. GS2E overcomes these limitations by first reconstructing photorealistic static scenes using 3D Gaussian Splatting, and subsequently employing a novel, physically-informed event simulation pipeline. This pipeline generally integrates adaptive trajectory interpolation with physically-consistent event contrast threshold modeling. Such an approach yields temporally dense and geometrically consistent event streams under diverse motion and lighting conditions, while ensuring strong alignment with underlying scene structures. Experimental results on event-based 3D reconstruction demonstrate GS2E's superior generalization capabilities and its practical value as a benchmark for advancing event vision research.

Figures (7)

• Figure 1: We propose GS2E, a high-fidelity synthetic dataset designed for 3D event-based vision, comprising over 1150 scenes. GS2E examples of RGB frames and event streams are shown above.
• Figure 2: Overview and comparison of event-based 3D dataset construction methods. We compare (1) real-world capture, (2) video-driven synthesis, and (3) simulation via computer graphics engines in terms of commonly used methods, strengths, and drawbacks.
• Figure 3: Overview of the proposed GS2E pipeline. Starting from sparse multi-view RGB images and known camera poses, we reconstruct high-fidelity scene representations using 3D Gaussian Splatting. Virtual camera trajectories are then synthesized via velocity-aware reparameterization and interpolation. The rendered image sequences are passed to a volumetric event simulator to generate temporally coherent and geometrically consistent event streams.
• Figure 4: Qualitative comparison of synthesized event distributions using GS2E versus traditional video-driven event synthesis methods, evaluated against real-world event data from the DSEC dataset.
• Figure 5: Application to Multiple Tasks. We benchmark it across event-vision tasks: 3D reconstruction and image deblurring.