E-3DGS: Gaussian Splatting with Exposure and Motion Events

Xiaoting Yin · Hao Shi · Yuhan Bao · Zhenshan Bing · Yiyi Liao · Kailun Yang · Kaiwei Wang

E-3DGS incorporate exposure event information into event-based 3D Gaussian Splatting (3DGS), converting sparse events during exposure into dense intensity frames for high-quality event-based 3D reconstruction. To achieve high-quality 3D reconstruction using only a single event sensor, we combine motion and exposure events to balance quality and efficiency in high-speed scenarios. In Fast Reconstruction Mode, E-3DGS achieves a PSNR gain of 5.68dB over EventNeRF, along with a significantly higher rendering speed (79.37 FPS vs. 0.03 FPS). In the HighQuality Reconstruction Mode, E-3DGS delivers a PSNR increase of 10.89 dB compared to the event-to-grayscale learning-based 3DGS baseline. In the future, we intend to explore the application and enhancement of E-3DGS based on exposure events for dense robotic perception, aiming to achieve higher-quality large-scale dense 3D reconstruction using purely event-based data.

TODO

• Release the code.
• Release the arxiv preprint.

Citation

If our work is helpful to you, please consider citing us by using the following BibTeX entry:

@misc{yin2024e3dgsgaussiansplattingexposure,
      title={E-3DGS: Gaussian Splatting with Exposure and Motion Events}, 
      author={Xiaoting Yin and Hao Shi and Yuhan Bao and Zhenshan Bing and Yiyi Liao and Kailun Yang and Kaiwei Wang},
      year={2024},
      eprint={2410.16995},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.16995}, 
}