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

Objects behind a diffuser are observed with varying blur depending on their location and the viewpoint, making the recovery of the clear scene challenging. Unlike defocus or motion blur, the blurring caused by a diffuser is a phenomenon that depends on the three-dimensional geometry of the scene, including the shape of the diffuser. This paper proposes a method to deblur the scene behind the diffuser from the observations from multiple viewpoints through inverse rendering. By representing the scene with 3D Gaussian primitives, the blurry images from an arbitrary viewpoint can be rendered efficiently. By optimizing to reproduce the observations from each viewpoint, we can obtain a Gaussian distribution representing a consistent, shared, clear scene, along with the blur parameter of the diffuser. Comparative experiments have been done to show that our method outperforms conventional techniques based on radiance fields and blind deconvolution. We also demonstrate that the proposed method improves the accuracy of the downstream task of text recognition.