Research map / 3D vision / Gaussian splatting
CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images
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 CoCoGaussian, a 3D Gaussian Splatting method that models the circle of confusion from depth and a learnable aperture to reconstruct sharp scenes from defocused multi-view images.
Problem
Reconstructing sharp 3D scenes from multi-view images affected by defocus blur due to finite depth of field, which existing 3DGS methods handle poorly without photographic defocus modeling.
Approach
Generates multiple CoC Gaussians per base Gaussian, with offsets set by a physically derived CoC diameter (from depth, learnable aperture K and focus plane) and a learnable scaling factor beta to handle unreliable depth; the rendered sets are combined with a per-pixel learned weighted sum.
Key results
- On the Deblur-NeRF synthetic dataset, CoCoGaussian reports 30.84 PSNR, 0.9212 SSIM and 0.0478 LPIPS, the best across all three metrics.
- On the DoF-NeRF real-world dataset, it reaches 30.14 PSNR and 0.0701 LPIPS, versus 29.87 PSNR and 0.1100 LPIPS for BAGS.
- On all-in-focus NeRF-LLFF, it improves over 3DGS (27.76 vs 27.10 PSNR, 0.0372 vs 0.0569 LPIPS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoCoGaussian instead of BAGS | 3D scene reconstruction from defocused images · Deblur-NeRF synthetic | PSNR: 30.65 -> 30.84 | Helped |
| Use CoCoGaussian instead of PDRF-10 | 3D scene reconstruction from defocused images · Deblur-NeRF real-world | PSNR: 23.85 -> 23.70 | Hurt |
| Use CoCoGaussian instead of BAGS | 3D scene reconstruction from defocused images · DoF-NeRF real-world | LPIPS: 0.1100 -> 0.0701 | Helped |
| Use CoCoGaussian instead of 3DGS | novel view synthesis on all-in-focus images · NeRF-LLFF | PSNR: 27.10 -> 27.76 | Helped |
| Add CoC modeling (ablation) | CoCoGaussian · DoF-NeRF real-world | PSNR: 28.29 -> 30.14 | Helped |
| Add learnable CoC direction vectors instead of fixed directions (ablation) | CoCoGaussian · DoF-NeRF real-world | PSNR: 28.91 -> 30.14 | Helped |
| Add the CoC scaling factor (ablation) | CoCoGaussian · DoF-NeRF real-world | PSNR: 29.46 -> 30.14 | Helped |
| Add aperture parameter K (ablation) | CoCoGaussian · DoF-NeRF real-world | PSNR: 27.37 -> 30.14 | Helped |
Surprising findings
- Without the CoC modeling, the ablated model (w/o CoC) still reaches 28.29 PSNR, and the learnable scaling factor alone is insufficient to match the full model, indicating physically grounded CoC contributes most of the gain.
- Deblurring 3DGS shows minimal improvement over vanilla 3DGS on the higher-resolution DoF-NeRF dataset, attributed to its limited kernel size.
Limitations
- The learnable CoC scaling factor is bounded, so it adapts mainly when the CoC diameter is overestimated, not underestimated.
- Real-world Deblur-NeRF dataset has illumination differences between sharp and defocused images, lowering PSNR.
- BAGS requires manual blur kernel size tuning, a limitation noted by the authors for comparison rather than their own method.
Experimental setup
Datasets: Deblur-NeRF synthetic dataset Deblur-NeRF real-world dataset DoF-NeRF real-world dataset NeRF-LLFF
Benchmarks: Deblur-NeRF synthetic Deblur-NeRF real-world DoF-NeRF real-world NeRF-LLFF
Metrics: PSNR SSIM LPIPS
Baselines: NeRF 3DGS Deblur-NeRF PDRF DP-NeRF Deblurring 3DGS BAGS
Builds on: 3D Gaussian Splatting Deblurring 3DGS BAGS DoF-NeRF Deblur-NeRF
Closest papers by meaning
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal (CVPR 2025 · similarity 0.90)
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (CVPR 2025 · similarity 0.90)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.90)
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.