Research map / 3D vision / Gaussian splatting
Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos
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 a unified dynamic Gaussian Splatting framework that handles both defocus and motion blur in monocular videos by predicting per-pixel blur kernels and intensity with a blur prediction network under a blur-aware sparsity constraint, plus dynamic Gaussian densification and unseen-view supervision.
Problem
Novel view synthesis of dynamic scenes from monocular videos degraded by either defocus blur or motion blur; existing dynamic Gaussian methods handle only one blur type and rely on sharp inputs.
Approach
Model both defocus and motion blur as per-pixel blur-kernel convolution blended with sharp rendering; predict kernels and blur intensity with a BP-Net conditioned on camera embedding, rendered image/depth/mask, and pixel coordinates; constrain kernel sparsity by blur intensity; densify dynamic Gaussians via foreground remapping; and supervise with unseen-view appearance warped from training views.
Key results
- On D2RF defocus blur, achieves 29.39 PSNR / 0.859 SSIM / 0.078 LPIPS, versus 28.49 PSNR for De4DGS.
- On DyBluRF motion blur, achieves 27.01 PSNR / 0.876 SSIM / 0.056 LPIPS, versus 26.62 PSNR for De4DGS.
- Renders at 65.143 FPS at 512x288 on an RTX 3090, with about 1 hour training per scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use unified blur synthesis with BP-Net and the proposed scene optimization | Dynamic Gaussian Splatting · D2RF defocus blur and DyBluRF motion blur | PSNR: D2RF: 29.39 vs 28.32; DyBluRF: 27.01 vs 26.62 | Helped |
| add blur-aware sparsity constraint (ablation) | Dynamic Gaussian Splatting · D2RF defocus blur | PSNR: 29.03 -> 29.39 | Helped |
| add blur-aware sparsity constraint (ablation) | Dynamic Gaussian Splatting · DyBluRF motion blur | PSNR: 26.63 -> 27.01 | Helped |
| add BP-Net shortcut connections (ablation) | BP-Net · D2RF defocus blur | PSNR: 29.12 -> 29.39 | Helped |
| add BP-Net shortcut connections (ablation) | BP-Net · DyBluRF motion blur | PSNR: 26.74 -> 27.01 | Helped |
| add dynamic Gaussian densification (ablation) | Dynamic Gaussian Splatting · D2RF defocus blur | PSNR: 29.19 -> 29.39 | Helped |
| add dynamic Gaussian densification (ablation) | Dynamic Gaussian Splatting · DyBluRF motion blur | PSNR: 26.53 -> 27.01 | Helped |
| use unseen view information during training (ablation) | Dynamic Gaussian Splatting · D2RF defocus blur | PSNR: 29.09 -> 29.39 | Helped |
| use unseen view information during training (ablation) | Dynamic Gaussian Splatting · DyBluRF motion blur | PSNR: 26.66 -> 27.01 | Helped |
| use blur kernel size K=11 rather than K=9 (ablation) | Dynamic Gaussian Splatting · D2RF defocus blur | PSNR: 28.29 -> 28.30 | No clear effect |
Surprising findings
- Feeding video-deblurred images to dynamic reconstruction baselines often yields only marginal gains, which the authors attribute to lack of 3D consistency in deblurred frames.
Limitations
- Depends on 2D priors (depth estimation, segmentation), whose errors degrade results.
- Fails on dynamic scenes with large non-rigid motion blur.
- Requires per-scene optimization like vanilla 3DGS.
Experimental setup
Datasets: D2RF DyBluRF D2RF-v2 DyBluRF-v2 Deblur-NeRF
Benchmarks: D2RF DyBluRF D2RF-v2 DyBluRF-v2 Deblur-NeRF
Metrics: PSNR SSIM LPIPS KL divergence FPS
Baselines: D3DGS Shape-of-Motion D2RF DyBluRF De4DGS DeformableGS BAGS De3DGS BSSTNet
Builds on: Shape-of-Motion 3D Gaussian Splatting D2RF DyBluRF Deblur4DGS
Closest papers by meaning
- BARD-GS: Blur-Aware Reconstruction of Dynamic Scenes via Gaussian Splatting (CVPR 2025 · similarity 0.91)
- MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors (ICLR 2025 · similarity 0.90)
- 4D Gaussian Splatting in the Wild with Uncertainty-Aware Regularization (NeurIPS 2024 · similarity 0.90)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · similarity 0.90)
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.