MoBGS: Motion Deblurring Dynamic 3D Gaussian Splatting for Blurry Monocular Video

TL;DR

MoBGS tackles dynamic novel view synthesis from blurry monocular video by explicitly modeling motion blur from both global camera motion and local object motion. It introduces BLCE, which guides latent camera pose estimation using frame blur, and LCEE, which estimates a latent exposure duration to unify blur across the scene, all within a dynamic 3D Gaussian Splatting framework. The method represents scenes with static and dynamic Gaussians and uses spline-based motion, with Neural ODEs enabling smooth, blur-aware pose trajectories. Extensive experiments on Stereo Blur and DAVIS demonstrate state-of-the-art perceptual quality and temporal consistency, alongside substantial speed advantages over recent methods.

Abstract

We present MoBGS, a novel motion deblurring 3D Gaussian Splatting (3DGS) framework capable of reconstructing sharp and high-quality novel spatio-temporal views from blurry monocular videos in an end-to-end manner. Existing dynamic novel view synthesis (NVS) methods are highly sensitive to motion blur in casually captured videos, resulting in significant degradation of rendering quality. While recent approaches address motion-blurred inputs for NVS, they primarily focus on static scene reconstruction and lack dedicated motion modeling for dynamic objects. To overcome these limitations, our MoBGS introduces a novel Blur-adaptive Latent Camera Estimation (BLCE) method using a proposed Blur-adaptive Neural Ordinary Differential Equation (ODE) solver for effective latent camera trajectory estimation, improving global camera motion deblurring. In addition, we propose a Latent Camera-induced Exposure Estimation (LCEE) method to ensure consistent deblurring of both a global camera and local object motions. Extensive experiments on the Stereo Blur dataset and real-world blurry videos show that our MoBGS significantly outperforms the very recent methods, achieving state-of-the-art performance for dynamic NVS under motion blur.

Figures (10)

• Figure 1: Our MoBGS achieves state-of-the-art dynamic deblurring novel view synthesis on (a) real-world casually captured blurry monocular videos ponttuset20182017davischallengevideo and (b) synthesized blurry monocular videos sun2024dyblurf, delivering (c) high perceptual quality (best LPIPS and MUSIQ scores) and fast rendering ($\sim$500 FPS). Each region is cropped and enlarged.
• Figure 2: Overview of MoBGS. MoBGS accurately models scene blurriness by jointly considering global camera and local object motion over the same exposure time. It first estimates latent camera poses for each blurry frame using the Blur-adaptive Latent Camera Estimation (BLCE) method. Then, leveraging these poses, it estimates the corresponding exposure time via the Latent Camera-induced Exposure Estimation (LCEE) method, ensuring a consistent blur modeling of local moving objects.
• Figure 3: Visual comparisons for dynamic deblurring novel view synthesis on the Stereo Blur dataset.
• Figure 4: Visual comparisons for dynamic deblurring novel view synthesis on the DAVIS dataset. 'Train.' refers to the training frame. The results of DyBluRF sun2024dyblurf exhibit significant misalignment, as its latent camera pose optimization is overfitted to the training poses. Full image results are provided in the Supplementary.
• Figure 5: Visual comparisons for BLCE ablation study.