Research map / 3D vision / Gaussian splatting

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

CVPR 2025 · 3D vision › Gaussian splatting · approach: Deformable 4D Gaussian splatting · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BARD-GS, a two-stage Gaussian Splatting method that separately models camera-induced and object-induced motion blur to reconstruct dynamic scenes from blurry monocular inputs, and releases a real-world blurry dynamic-scene dataset.

Problem

Reconstructing dynamic scenes from blurry monocular images with imprecise camera poses, where camera motion and object motion both cause blur that existing dynamic 3DGS and NeRF methods handle poorly.

Approach

Decompose motion blur into camera motion blur and object motion blur and model each explicitly: first optimize virtual camera poses within exposure to deblur static regions, then learn a time-conditioned deformation field with virtual timestamps per exposure to deblur dynamic regions, with a 3D track loss from BootsTAPIR and Depth-Anything to constrain Gaussian trajectories.

Key results

• On the real-world blurry dataset, BARD-GS reaches 25.13 SI-PSNR and 0.1567 SI-LPIPS for novel view synthesis, versus 22.83 SI-PSNR for D3DGS (best baseline by SI-PSNR was D3DGS+MPRNet at 23.10).
• On the synthesized deblurring benchmark, BARD-GS achieves 28.75 PSNR and 0.2343 LPIPS, outperforming 4DGS+MPRNet (28.17 PSNR).
• Ablation shows removing static deblurring causes the largest drop (deblur PSNR 26.00 vs 28.75).

Measured effects reported in the paper

Surprising findings

• Removing dynamic deblurring has less impact on the synthesized dataset than on the real dataset, because synthetic scenes are less blurred.
• Some deblurred outputs appear sharper than the ground truth in DyCheck, since the GT was captured at a low frame rate and is itself blurry.

Limitations

• Real-world dataset has only 12 scenes, each with two GoPros; evaluation is limited to novel view synthesis on it.
• Increasing virtual views helps only up to a point, after which performance degrades, so the number must be tuned.
• Relies on off-the-shelf components (COLMAP, TrackAnything, BootsTAPIR, Depth-Anything) whose errors propagate.

Experimental setup

Datasets: HyperNeRF DyCheck BARD-GS real-world blurry dataset

Benchmarks: HyperNeRF blurry synthetic subset DyCheck blurry synthetic subset BARD-GS real-world blurry dataset

Metrics: PSNR SSIM LPIPS Laplacian Variance MUSIQ SI-PSNR SI-SSIM SI-LPIPS

Baselines: DyBluRF D3DGS 4DGaussians MPRNet + D3DGS MPRNet + 4DGaussians

Builds on: 3D Gaussian Splatting BAD-Gaussians Deformable-3D-Gaussians 4D Gaussian Splatting BootsTAPIR

Closest papers by meaning

• GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring (NeurIPS 2024 · similarity 0.91)
• Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.91)
• Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos (NeurIPS 2025 · similarity 0.91)
• MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
• HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
• MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors (ICLR 2025 · similarity 0.89)