BeSplat: Gaussian Splatting from a Single Blurry Image and Event Stream

Gopi Raju Matta^*    Reddypalli Trisha   Kaushik Mitra  

---

This paper was accepted by the 1st Workshop on "Event-based Vision in the Era of Generative AI - Transforming Perception and Visual Innovation", held at WACV 2025.

If you like our project, please give us a star ⭐ on GitHub.

We explore the possibility of recovering sharp radiance fields (Gaussian splats) and camera motion trajectory from a single motion-blurred image. This allows BeSplat to decode the underlying sharp scene representation and video from a single blurred image and its corresponding event stream.

🚀 Updates

• Project homepage is now live! Check it out here.
• Training, testing, and evaluation codes, along with datasets, are now available.

🔍 Approach

Given a single motion-blurred image and its corresponding event stream, BeSplat recovers the underlying 3D scene representation (Gaussian splats) and the camera motion trajectory jointly. Specifically, we represent the 3D scene using Gaussian Splatting and model the camera motion trajectory with a Bézier curve in SE(3) space. Both the blurry image and the accumulated events over a time interval can be synthesized from the 3D scene representation using the estimated camera poses. The scene representation and camera trajectory are optimized by minimizing the discrepancy between the synthesized data and the real-world measurements.

🛠️ Setup Instructions

Installation

Follow the setup instructions for 3D Gaussian Splatting for environment requirements and setup.

Download Datasets

We use real-world datasets from E2NeRF, captured using the DAVIS346 color event camera, and synthetic datasets from BeNeRF for evaluations.

• The real-world datasets contain five scenes: letter, lego, camera, plant, and toys.
• The synthetic datasets from BeNeRF include three sequences from Unreal Engine: livingroom, whiteroom, and pinkcastle, and two sequences from Blender: tanabata and outdoorpool.

You can download the datasets from the following links:

• Download BeNeRF Datasets

Training

🔹 Real Dataset Training

python train_modified_final_real.py \
  -s <source_path>

🔹 Synthetic Dataset Training

python train_modified_final_synth.py \
  -s <source_path>

Command Line Arguments:

• index: Index of the input sequence to process
• rgb_coeff: Weight for RGB loss component
• event_coeff_real: Weight for event-based loss on real datasets
• blur_sample_num: number of key frames for trajectory time sampling
• deblur: switch the deblur mode
• mode: models of camera motion trajectory (i.e. Linear, Spline, Bezier)
• bezier_order: order of the Bézier curve when use Bézier curve for trajectory modeling

📌 Note:

We used rgb_coeff=0.1 and event_coeff_real=1.0 for real datasets, and rgb_coeff=1.0 and event_coeff_real=2.0 for synthetic datasets.
Please try out different values to get better results for your custom dataset.

🧪 Evaluation

Rendering

python render.py -m <path to trained model>  # Generate renderings

Additional Command Line Arguments for render.py:

• optim_pose: optimize the camera pose to align with the dataset

Render Video

python render_video.py -m <path to trained model>

Results

You can check our results at the following link.

• https://gopirajumatta.github.io/BeSplat/

✒️ Citation

If you find this repository useful, please consider citing our paper:

@InProceedings{Matta_2025_WACV,
    author    = {Matta, Gopi Raju and Reddypalli, Trisha and Mitra, Kaushik},
    title     = {BeSplat: Gaussian Splatting from a Single Blurry Image and Event Stream},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops},
    month     = {February},
    year      = {2025},
    pages     = {917-927}
}

🙏 Acknowledgment

In our work, the camera trajectory optimization was inspired by Deblur-GS, and the event stream integration into Gaussian Splatting was inspired by the methodology used in BeNeRF. The overall code framework is based on both repositories. We appreciate the efforts of the contributors to these amazing projects.