Geometry-aware deblurring of blur caused by a diffuser using 3D Gaussian splatting.

Area of Science:

• Computer Vision
• Computational Imaging
• 3D Scene Reconstruction

Background:

• Diffuser-induced blur is viewpoint-dependent and complex, unlike simpler blur types.
• Recovering clear scenes from diffuser-blurred images is challenging due to 3D scene geometry and diffuser shape.
• Existing methods struggle with the unique challenges of diffuser blur.

Purpose of the Study:

• To develop an effective method for deblurring scenes viewed through a diffuser.
• To enable accurate 3D scene reconstruction from multiple blurred viewpoints.
• To improve performance on downstream tasks like text recognition.

Main Methods:

• Utilizes inverse rendering to deblur scenes behind a diffuser.
• Represents the scene using 3D Gaussian primitives for efficient rendering.
• Optimizes a shared Gaussian distribution to match observations from multiple viewpoints.

Main Results:

• Successfully reconstructs a consistent, clear 3D scene and diffuser blur parameters.
• Outperforms conventional radiance field and blind deconvolution techniques.
• Demonstrates improved accuracy in text recognition tasks.

Conclusions:

• The proposed inverse rendering method effectively handles diffuser blur.
• 3D Gaussian primitives enable efficient and accurate scene deblurring.
• This approach offers a significant advancement for computational imaging and scene reconstruction.