Abstract
In this paper, we propose the first Structure-from-Motion (SfM)-free deblurring 3D Gaussian Splatting method via event camera, dubbed DeblurSplat. We address the motion-deblurring problem in two ways. First, we leverage the pretrained capability of the dense stereo module (DUSt3R) to directly obtain accurate initial point clouds from blurred images. Without calculating camera poses as an intermediate result, we avoid the cumulative errors transfer from inaccurate camera poses to the initial point clouds' positions. Second, we introduce the event stream into the deblur pipeline for its high sensitivity to dynamic change. By decoding the latent sharp images from the event stream and blurred images, we can provide a fine-grained supervision signal for scene reconstruction optimization. Extensive experiments across a range of scenes demonstrate that DeblurSplat not only excels in generating high-fidelity novel views but also achieves significant rendering efficiency compared to the SOTAs in deblur 3D-GS.
Introduction
3D reconstruction techniques like 3D Gaussian Splatting (3D-GS) have revolutionized novel view synthesis. However, they are highly sensitive to input quality. In real-world scenarios, fast camera movement or low light often results in motion blur, which causes two critical failures:
- Initialization Failure: Traditional pipelines rely on Structure-from-Motion (SfM) tools like COLMAP. Blurry images impede feature matching, leading to sparse or failed initializations.
- Optimization Ambiguity: A blurry image is a temporal average, making the recovery of a "latent sharp" scene mathematically underdetermined.
DeblurSplat is the first SfM-free 3D-GS framework designed to overcome these hurdles by integrating a pretrained dense stereo module and high-speed event camera data.
Methodology
SfM-free Initialization with DUSt3R
Traditional deblurring methods attempt to estimate camera poses from blurry images using COLMAP, which often fails or produces inaccurate "coarse" point clouds. DeblurSplat replaces this with DUSt3R, a transformer-based dense stereo model. DUSt3R directly estimates 3D point maps and relative poses from image pairs, providing a robust starting point even under severe blur.
Confidence Balanced Sampling (CBS)
While DUSt3R provides point clouds, it also outputs confidence scores. Simply filtering for high-confidence points can lose the structural "skeleton" of the scene. DeblurSplat introduces Confidence Balanced Sampling (CBS), which partitions points by confidence levels and samples them to ensure both high-quality detail and complete spatial coverage.
Progressive Alignment via Event Cameras
Event cameras capture per-pixel brightness changes at microsecond resolution, making them immune to motion blur. DeblurSplat utilizes an Event-based Double Integral (EDI) model to decode blurred RGB images into a sequence of latent sharp images. These sharp images serve as a fine-grained supervision signal, guiding the optimization of 3D Gaussians and iteratively refining camera poses.
Experiments and Results
The method was evaluated on synthetic datasets (Normal and Hard) and real-world data captured with a DAVIS346 event camera.
- Quantitative Performance: In challenging synthetic scenes with large-scale camera movements, DeblurSplat achieved a +2.74 dB increase in PSNR over state-of-the-art methods like BAD-GS and EDNeRF.
- Real-World Robustness: In severe blur scenarios, DeblurSplat maintained a PSNR of 27.98 dB, significantly outperforming the next best method (21.53 dB).
- Efficiency: By bypassing traditional SfM, initialization is approximately 3x faster. The final model supports high-speed rendering at over 210 FPS.
Conclusion
DeblurSplat demonstrates that for robust deblurring, the initialization stage is as vital as the rendering optimization. By leveraging DUSt3R for SfM-free initialization and event cameras for temporal priors, the framework achieves superior reconstruction quality and efficiency. Future work will address rolling shutter effects and scaling to larger environments.