BSGS: Bi-stage 3D Gaussian Splatting for Camera Motion Deblurring

TL;DR

This work tackles reconstructing sharp 3D scenes from motion-blurred images by introducing Bi-stage 3D Gaussian Splatting (BSGS), a two-stage framework that first refines camera poses and then applies a global rigid transformation to Gaussian primitives. It introduces subframe gradient aggregation to resolve multi-view gradient conflicts and Space-Time Coupling Densification to adaptively control Gaussian densification during training, preventing premature noisy splitting in blurred regions. The method jointly optimizes photometric consistency between blurred inputs and synthetic renderings within a differentiable 3D Gaussian splatting pipeline, achieving state-of-the-art deblurring performance and robust novel-view synthesis on real and synthetic data. The approach reduces reliance on precise pose estimates and demonstrates practical efficiency by using fewer primitives while maintaining high-quality reconstructions, highlighting its potential for real-world 3D reconstruction under challenging blur conditions.

Abstract

3D Gaussian Splatting has exhibited remarkable capabilities in 3D scene reconstruction. However, reconstructing high-quality 3D scenes from motion-blurred images caused by camera motion poses a significant challenge.The performance of existing 3DGS-based deblurring methods are limited due to their inherent mechanisms, such as extreme dependence on the accuracy of camera poses and inability to effectively control erroneous Gaussian primitives densification caused by motion blur. To solve these problems, we introduce a novel framework, Bi-Stage 3D Gaussian Splatting, to accurately reconstruct 3D scenes from motion-blurred images. BSGS contains two stages. First, Camera Pose Refinement roughly optimizes camera poses to reduce motion-induced distortions. Second, with fixed rough camera poses, Global RigidTransformation further corrects motion-induced blur distortions. To alleviate multi-subframe gradient conflicts, we propose a subframe gradient aggregation strategy to optimize both stages. Furthermore, a space-time bi-stage optimization strategy is introduced to dynamically adjust primitive densification thresholds and prevent premature noisy Gaussian generation in blurred regions. Comprehensive experiments verify the effectiveness of our proposed deblurring method and show its superiority over the state of the arts.Our source code is available at https://github.com/wsxujm/bsgs

Figures (6)

• Figure 1: The pipeline of our method. We propose a Bi-stage approach for robust 3D reconstruction from motion-blurred images. In the first stage, we refine the camera poses to obtain an approximate estimate. In the second stage, we fix the optimized camera poses and apply global rigid transformations to the Gaussian points, simulating the motion blur. We optimize the Gaussian scene by minimizing photometric errors and introduce a subframe gradient aggregation strategy to resolve conflicts.
• Figure 2: Subframe Gradient Aggregation via Max-Pooling.
• Figure 3: Visualization of Space-Time Coupling Densification effects. a) The sharp 2D image to be reconstructed; b) The initialized point cloud; c) The point cloud obtained after applying the time-aware threshold $\tau_t$ and the space-aware threshold $\tau_s$ in Stage I, with the general scene outline and low-frequency foreground been roughly reconstructed; d) The final result after applying $\tau_t$ and $\tau_s$ in Stage II. The results demonstrate that Space-Time Coupling Densification can effectively control erroneous Gaussian primitives densification caused by motion blur and improve the reconstruction performance of detail information.
• Figure 4: Qualitative comparison for deblurring on the synthetic and real-world scenes. The first two rows are from synthetic datasets while the last two rows are from real-world datasets. The results demonstrate our method's superior deblurring and detail restoration performance compared to prior approaches.
• Figure 5: Visual comparison with error maps in Trolley scene.