SweepEvGS: Event-Based 3D Gaussian Splatting for Macro and Micro Radiance Field Rendering from a Single Sweep

Jingqian Wu, Shuo Zhu, Chutian Wang, Boxin Shi, Edmund Y. Lam

Abstract

Reconstructing high-quality 3D scenes typically requires a dense set of sharp images, a requirement that becomes increasingly difficult to meet during rapid camera motion or in specialized macro and micro imaging contexts. Motion blur often compromises the fidelity of traditional 3D Gaussian Splatting (3D-GS) and Neural Radiance Fields (NeRF). We present SweepEvGS, a hardware-integrated pipeline that leverages the high temporal resolution of event cameras to supervise a 3D-GS model from a single camera sweep. By utilizing direct event supervision and a specialized noise-filtering process, SweepEvGS achieves sharp reconstructions even under significant motion. Our method demonstrates a 150x speedup in rendering compared to NeRF-based event models and requires 3x less GPU memory, establishing a new benchmark for efficient, scale-invariant 3D reconstruction.

Introduction: The Challenge of Motion Blur

Creating high-quality Radiance Fields usually necessitates a "dense" set of sharp, high-resolution images. In practice, this capture process is often fraught with difficulty. Rapid camera movement intended to capture a scene quickly inevitably introduces motion blur. Conversely, moving the camera slowly enough to maintain image sharpness makes the process tedious and inefficient.

This problem is particularly acute in macro and micro imaging. At these scales, even microscopic vibrations or slight camera movements result in significant blur, rendering traditional 3D-GS or NeRF techniques ineffective. While event cameras—sensors that record changes in brightness at microsecond resolution—are inherently immune to motion blur, existing event-based reconstruction methods (such as EventNeRF) are computationally expensive and often struggle with real-world sensor noise.

Methodology: The SweepEvGS Pipeline

SweepEvGS introduces a hardware-integrated pipeline designed to build a 3D scene from a single "sweep" of an event camera.

The SweepEvGS pipeline: (a) A single camera sweep captures (b) event signals and an initial RGB frame to supervise (c) the 3D Gaussian Splatting model.

The Core Workflow

  1. Initial Anchor: The system captures a single static RGB frame at the start of the sweep to provide a clean baseline and assist with initial camera calibration.
  2. Event Supervision: As the camera sweeps across the scene, it records a dense stream of events. Rather than reconstructing intermediate frames, SweepEvGS uses these events directly to supervise the 3D-GS model.
  3. Linlog Mapping: Because event cameras respond to logarithmic changes in brightness, the authors employ a "linlog" mapping. This allows the model to accurately compare predicted brightness changes from the 3D Gaussians against the ground-truth event stream.
  4. Noise Filtering: Real-world event data is inherently noisy, especially in microscopic settings. The researchers implemented a specialized $y$-noise filter to strip out interference signals before training, ensuring the model learns the scene structure rather than sensor artifacts.

Experimental Results

The paper validates SweepEvGS across synthetic sequences and real-world macro and micro environments. The micro-imaging benchmark is particularly notable, featuring high-resolution event data of biological specimens like onion skin and flies captured via a widefield microscope.

Qualitative comparison between standard 3D-GS and SweepEvGS in macro and micro settings. Standard 3D-GS suffers from severe motion blur, while SweepEvGS recovers sharp details.

Key Insights and Performance

  • Events as the Primary Signal: SweepEvGS demonstrates that the differential signal provided by events is superior for reconstruction during fast motion, as it preserves high-frequency details lost to blur in standard frames.
  • Efficiency Gains: By utilizing 3D Gaussian Splatting instead of NeRF-based architectures, the authors bypassed expensive volumetric rendering. This results in rendering speeds of approximately 60 FPS—a 150x speedup over previous event-based NeRFs.
  • Resource Efficiency: The method requires roughly 3x less GPU memory (5GB vs 15GB) during the rendering stage and reduces training time from 14 hours (for NeRF-based models) to just 8–10 minutes per scene.
  • Scale Invariance: The system's robustness is evident across scales, successfully reconstructing intricate textures on cloth (macro) and fine structures on a fly's wing (micro) using the same underlying logic.

Conclusion and Future Work

SweepEvGS proves that event-based 3D-GS is a generalized, highly efficient solution for sweep-based scanning. Currently, the system relies on precise camera poses provided by industrial motion controllers. Future work aims to integrate event-based visual odometry, which would enable handheld 3D scanning by simply sweeping an event camera over an object, making high-fidelity 3D reconstruction truly portable.

Summary updated

BibTeX
@article{wu2025sweepevgs,
  author = {Jingqian Wu and Shuo Zhu and Chutian Wang and Boxin Shi and Edmund Y. Lam},
  title = {{SweepEvGS: Event-Based 3D Gaussian Splatting for Macro and Micro Radiance Field Rendering from a Single Sweep}},
  journal = {IEEE Trans. Circuits Syst. Video Technol.},
  year = {2025},
  doi = {10.1109/tcsvt.2025.3583735},
}
Figure