Research map / 3D vision / Gaussian splatting
DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting
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.
Introduces DoF-Gaussian, a lens-based 3D Gaussian Splatting method that learns aperture and focus parameters from shallow-DoF multi-view images to render sharp novel views and controllable depth-of-field effects.
Problem
3D Gaussian Splatting assumes pinhole, all-in-focus inputs, so shallow depth-of-field images degrade reconstruction and depth estimation, and existing DoF/deblurring datasets do not evaluate refocusing.
Approach
Replace the pinhole camera with a differentiable lens model whose aperture and focus distance are learnable per image, combine it with per-scene fine-tuned monocular depth priors, and apply a defocus-to-focus reweighting schedule after lens parameters converge.
Key results
- On the Deblur-NeRF defocus dataset, the method reports best average metrics versus DoF-NeRF, DP-NeRF, BAGS and Deblurring 3DGS (exact averages not in extracted text).
- On the authors' synthetic dataset, PSNR 28.70 / SSIM 0.864 / LPIPS 0.095 versus DoF-NeRF 25.59 / 0.788 / 0.207.
- In all-in-focus settings, PSNR 27.81 vs Mip-Splatting 27.05 on average over Real Forward-facing and T&T DB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the lens-based imaging model (ablation) | DoF-Gaussian · Ablation study | PSNR: 21.31 -> 23.05 | Helped |
| add defocus-to-focus adaptation (ablation) | DoF-Gaussian with lens model · Ablation study | PSNR: 23.05 -> 23.59 | Helped |
| add per-scene depth priors (ablation) | DoF-Gaussian with lens model · Ablation study | PSNR: 23.05 -> 23.42 | Helped |
| combine the lens model, defocus-to-focus adaptation, and per-scene depth priors (ablation) | DoF-Gaussian · Ablation study | PSNR: 21.31 -> 23.97 | Helped |
| use per-scene fine-tuned depth priors instead of depth without fine-tuning (ablation) | DoF-Gaussian · Depth-prior ablation | PSNR: 23.53 -> 23.97 | Helped |
| use per-scene fine-tuned depth priors instead of sparse depth maps (ablation) | DoF-Gaussian · Depth-prior ablation | PSNR: 23.44 -> 23.97 | Helped |
| use DoF-Gaussian instead of DoF-NeRF | Novel-view rendering with shallow-DoF inputs · Synthetic dataset | PSNR: 25.59 -> 28.70 | Helped |
| use DoF-Gaussian instead of Mip-Splatting | All-in-focus novel-view synthesis · Real Forward-facing and T&T DB | PSNR: 27.05 -> 27.81 | Helped |
Limitations
- Depth priors rely on COLMAP sparse reconstruction and a fine-tuned monocular depth network, adding pipeline complexity.
- Ideal geometric-optics CoC model differs from real DSLR CoC, which the defocus-to-focus adaptation only partially addresses.
Experimental setup
Datasets: Deblur-NeRF defocus blur dataset Real Forward-facing (LLFF) Tanks and Temples DB synthetic DoF dataset built from Real Forward-facing and T&T DB
Benchmarks: Deblur-NeRF defocus blur dataset Real Forward-facing Tanks and Temples DB DoF-Gaussian synthetic dataset
Metrics: PSNR SSIM LPIPS aperture error focus distance error
Baselines: Deblur-NeRF DoF-NeRF DP-NeRF BAGS Deblurring 3DGS Mip-Splatting
Builds on: 3D Gaussian Splatting Mip-Splatting Rade-GS DoF-NeRF Huang et al. 2DGS normal consistency
Closest papers by meaning
- DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal (CVPR 2025 · similarity 0.91)
- GES : Generalized Exponential Splatting for Efficient Radiance Field Rendering (CVPR 2024 · similarity 0.89)
- CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images (CVPR 2025 · similarity 0.89)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.89)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting (NeurIPS 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.