Abstract
Recent advances in 3D Gaussian Splatting (3DGS) have enabled high-fidelity scene reconstruction and real-time rendering. However, 3DGS remains highly sensitive to input quality, particularly in the presence of camera motion blur. Motion blur introduces two primary challenges: inaccurate camera pose estimation and erroneous densification, where the algorithm generates excessive "floater" primitives to compensate for image blur. We present BSGS (Bi-stage 3D Gaussian Splatting), a framework designed to reconstruct sharp 3D scenes from blurred 2D inputs. By utilizing a bi-stage optimization process, a novel gradient aggregation strategy, and space-time coupling densification, BSGS achieves state-of-the-art reconstruction quality while significantly reducing the number of Gaussian primitives required.
1. Introduction
3D Gaussian Splatting (3DGS) has emerged as a powerful alternative to Neural Radiance Fields (NeRF), offering faster training and superior rendering speeds. Despite its success, the "garbage in, garbage out" principle applies: 3DGS relies on sharp images and precise camera poses typically obtained via Structure-from-Motion (SfM). In real-world scenarios, camera jitter during exposure results in motion blur, which degrades SfM accuracy and causes 3DGS to fail. Specifically, the standard densification process in 3DGS mistakenly interprets blur as high-frequency detail, leading to the creation of noisy, memory-intensive "floaters" that degrade visual quality.
2. The BSGS Framework
BSGS addresses these challenges through three core technical innovations designed to stabilize optimization and refine scene geometry.
2.1 Bi-Stage Optimization
To decouple the errors introduced by motion blur, BSGS employs a two-stage optimization pipeline:
- Stage I: Camera Pose Refinement: The model focuses on optimizing the camera's trajectory during the exposure period. By modeling the motion path, the framework achieves a preliminary geometric alignment that accounts for the physical cause of the blur.
- Stage II: Global Rigid Transformation: Once the camera poses are stabilized, they are frozen. The model then applies a global rigid transformation to the Gaussian primitives themselves. This step corrects residual distortions that cannot be resolved through pose refinement alone, ensuring the final reconstructed geometry is sharp and coherent.
2.2 Subframe Gradient Aggregation via Max-Pooling
To simulate the blurring process, 3DGS methods often render multiple "subframes" and average them to match the blurred input. However, standard gradient averaging can lead to "gradient cancellation," where opposing signals from different subframes neutralize each other, stalling optimization. BSGS introduces Max-Pooling Gradient Aggregation. By selecting the strongest and most consistent gradient direction across subframes, the model prioritizes the most informative signals—effectively focusing on the "sharpest" moments within a blurred sequence to guide the optimization.
2.3 Space-Time Coupling Densification (STDC)
Standard 3DGS densification splits Gaussians based on a fixed view-space positional gradient threshold. In blurred scenes, this leads to premature and erroneous splitting. BSGS introduces Space-Time Coupling Densification (STDC), a dynamic thresholding strategy:
- Space-aware: It applies higher splitting thresholds to near-field objects, which are more susceptible to significant blur, preventing the accumulation of noise in the foreground.
- Time-aware: Thresholds are kept high during the initial pose-refinement stage to prevent the creation of floaters before the geometry is aligned. As the optimization progresses into the second stage, the thresholds are lowered to allow for the recovery of fine details.
3. Experimental Evaluation
3.1 Datasets and Protocol
The framework was evaluated on the Real-motion-blur dataset and the ExBluRF dataset (both synthetic and real-world), which features complex 6-DOF camera trajectories. To simulate realistic conditions, SfM was performed using only blurry images, providing the model with intentionally inaccurate initial poses and sparse point clouds.
3.2 Quantitative and Qualitative Results
BSGS achieved state-of-the-art performance across all benchmarks. On the ExBluRF real-world dataset, BSGS reached a PSNR of 29.16, outperforming previous leading methods such as BAD-GS (27.27) and Deblur-GS (25.35). Qualitatively, BSGS successfully recovers sharp textures, such as legible text and fine foliage edges, where competing methods produce significant smearing or artifacts.
3.3 Efficiency and Lean Reconstruction
A standout feature of BSGS is its efficiency. By implementing the "patience principle" through STDC—waiting for stable geometry before densifying—the model avoids unnecessary primitive creation. In the "Camellia" scene, BSGS utilized only 13,446 primitives, a massive reduction compared to the 225,000+ primitives required by BAD-GS. This results in a significantly lower memory footprint and faster rendering speeds without sacrificing quality.
4. Conclusion and Future Work
BSGS demonstrates that high-quality 3D reconstruction from blurred images is possible by carefully managing the densification process and optimizing camera trajectories. By prioritizing dominant gradient signals and employing a bi-stage refinement strategy, the framework produces sharper results with a fraction of the memory required by traditional 3DGS. While highly effective for camera motion blur, future work remains to address object motion blur and extremely high-frequency camera shakes that may exceed the current physical motion models.
Summary updated