BARD-GS: Blur-Aware Reconstruction of Dynamic Scenes via Gaussian Splatting

TL;DR

BARD-GS addresses dynamic scene reconstruction under motion blur by explicitly decoupling and modeling blur sources as camera- and object-induced. The method uses two sequential stages: camera motion deblurring via virtual camera trajectories and optimization of static Gaussians, followed by object motion deblurring with a time-conditioned deformation field to track dynamic Gaussians. A real-world motion blur dataset for dynamic scenes is introduced, enabling evaluation beyond synthetic benchmarks. Experimental results demonstrate improved rendering quality in dynamic regions and superior novel view synthesis compared with state-of-the-art baselines, supported by comprehensive ablations and metrics that emphasize sharpness and detail retention.

Abstract

3D Gaussian Splatting (3DGS) has shown remarkable potential for static scene reconstruction, and recent advancements have extended its application to dynamic scenes. However, the quality of reconstructions depends heavily on high-quality input images and precise camera poses, which are not that trivial to fulfill in real-world scenarios. Capturing dynamic scenes with handheld monocular cameras, for instance, typically involves simultaneous movement of both the camera and objects within a single exposure. This combined motion frequently results in image blur that existing methods cannot adequately handle. To address these challenges, we introduce BARD-GS, a novel approach for robust dynamic scene reconstruction that effectively handles blurry inputs and imprecise camera poses. Our method comprises two main components: 1) camera motion deblurring and 2) object motion deblurring. By explicitly decomposing motion blur into camera motion blur and object motion blur and modeling them separately, we achieve significantly improved rendering results in dynamic regions. In addition, we collect a real-world motion blur dataset of dynamic scenes to evaluate our approach. Extensive experiments demonstrate that BARD-GS effectively reconstructs high-quality dynamic scenes under realistic conditions, significantly outperforming existing methods.

Figures (10)

• Figure 1: The formation process of motion blur. It originates from two sources: camera-induced blur caused by camera movements during exposure, and object-induced blur resulting from fast objects moving. The static regions of a scene are affected solely by camera motion blur, while dynamic regions are impacted by both camera and object motion blur.
• Figure 2: An overview of the pipeline. Our method consists of two stages: camera motion deblur and object motion deblur. In the first stage, we handle camera motion blur by modeling the camera’s trajectory during each exposure, resulting in sharp reconstruction in the static regions. Then, we utilized the optimized camera poses together with the depth map obtain from DepthAnything to initialize the dynamic Gaussians. In the second stage, we address object motion blur by modeling the trajectory of 3D Gaussians within each exposure using deformation field, which allow us to achieve clear reconstruction in the dynamic regions.
• Figure 3: Demonstration of artifacts in synthesized dataset.
• Figure 4: Qualitative comparison of deblurring. The green boxes represent details in static region and the dashed yellow boxes denote the dynamic part. As Dycheck dataset is not captured at high frame rate, our method produces a better result than GT in the last row.
• Figure 5: Qualitative comparison of Novel View Synthesis. Our method shows a significantly better rendering quality in the dynamic region, e.g. the windmill, the face details, card and the toy car. Besides, for static region, our method reconstructs finer details and shows a even better result than the ground truth image, e.g. the details in the green bounding box in the second column.