Research map / 3D vision / Gaussian splatting

DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal

CVPR 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 7 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.

Extends 3D Gaussian Splatting with a thin-lens finite-aperture model and differentiable CoC-based per-Gaussian defocus rendering, enabling post-capture refocusing, aperture control, and all-in-focus rendering from defocused multi-view images.

Problem

3DGS assumes a pinhole camera and lacks explicit defocus modeling, so it cannot offer post-capture control of focal distance and aperture or handle defocus blur in multi-view training inputs.

Approach

Models each Gaussian's defocus by convolving it with an isotropic kernel sized by its thin-lens circle of confusion, makes the whole pipeline differentiable so per-view focal distance and aperture are learned, and adds a CoC-guided in-focus localization network that supervises all-in-focus renderings.

Key results

• On the real defocus dataset, DOF-GS reaches 24.12 PSNR / 0.7667 SSIM for all-in-focus novel view synthesis, comparable to Deblur-GS (24.12 / 0.7667) and BAGS (24.21 / 0.7631) and above Mip-Splatting (21.23 PSNR).
• On the synthetic dataset, DOF-GS reaches 29.39 PSNR / 0.8914 SSIM, close to the best baseline PDRF (29.96 / 0.8923).
• At 2400x1600 resolution, DOF-GS uses about 39 GB peak GPU memory versus BAGS which ran out of memory; at 1200x800 it uses 13.06 GB versus BAGS's 21.69 GB.

Measured effects reported in the paper

Surprising findings

• The two-parameter global camera model gives comparable, not superior, all-in-focus quality to per-pixel kernel deblurring methods, which the authors attribute to stricter constraints.

Limitations

• Regions that remain blurry in all input views cannot be recovered sharply.
• Camera poses are estimated by COLMAP and not optimized during training.

Experimental setup

Datasets: Deblur-NeRF defocus dataset (real and synthetic scenes)

Benchmarks: Deblur-NeRF real defocus dataset Deblur-NeRF synthetic defocus dataset

Metrics: PSNR SSIM GPU memory usage

Baselines: NeRF Mip-Splatting Deblur-NeRF DP-NeRF PDRF Deblur-GS BAGS

Builds on: 3D Gaussian Splatting Mip-Splatting surface splatting DP-NeRF Deblur-NeRF

Closest papers by meaning

• DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
• CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images (CVPR 2025 · similarity 0.90)
• GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.89)
• SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (CVPR 2025 · similarity 0.89)
• PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
• RestorGS: Depth-aware Gaussian Splatting for Efficient 3D Scene Restoration (CVPR 2025 · similarity 0.88)