Exploiting Deblurring Networks for Radiance Fields

Haeyun Choi1, Heemin Yang2, Janghyuk Han2, Sunghyun Cho2
1 KT, 2 POSTECH
CVPR 2025
arXiv Paper Supplementary Code Data

DeepDeblurRF is a novel deblurring radiance field approach that leverages DNN-based deblurring modules for efficient training and high-quality novel-view synthesis from blurry images, even with nonlinear outliers like noise and saturated pixels.

Teaser Image

We demonstrate the extensibility of our framework by constructing radiance fields using different scene representations, such as voxel grids and Gaussian Splatting. DeepDeblurRF-P and DeepDeblurRF-G are the results of our framework, where radiance fields are constructed using Plenoxels and 3D Gaussian Splatting, respectively.

Abstract

In this paper, we propose DeepDeblurRF, a novel radiance field deblurring approach that can synthesize high-quality novel views from blurred training views with significantly reduced training time. DeepDeblurRF leverages deep neural network (DNN)-based deblurring modules to enjoy their deblurring performance and computational efficiency. To effectively combine DNN-based deblurring and radiance field construction, we propose a novel radiance field (RF)guided deblurring and an iterative framework that performs RF-guided deblurring and radiance field construction in an alternating manner. Moreover, DeepDeblurRF is compatible with various scene representations, such as voxel grids and 3D Gaussians, expanding its applicability. We also present BlurRF-Synth, the first large-scale synthetic dataset for training radiance field deblurring frameworks. We conduct extensive experiments on both camera motion blur and defocus blur, demonstrating that DeepDeblurRF achieves state-of-the-art novel-view synthesis quality with significantly reduced training time.

DeepDeblurRF

DeepDeblurRF takes a set of blurred images of a scene and estimates a sharp radiance field that can synthesize a sharp novel view given an arbitrary pose.

Framework Image

To do so, DeepDeblurRF performs initial deblurring on the input blurred images. Then, it iteratively performs RF construction using deblurred images and RF-guided deblurring to gradually enhance the quality of the radiance field and the deblurred images. At the last iteration, we perform only the radiance field construction step and obtain a final radiance field, from which we can synthesize sharp novel views.

Iteration Image

As iterations progress, the rendered images contain high-quality scene information, which subsequently improves the RF-guided deblurring network's performance in the next iteration.

BlurRF Dataset

The BlurRF-Synth dataset is the first large-scale, multi-view dataset designed for radiance field deblurring approaches, capturing real-world camera degradations such as noise, saturated pixels, and in-camera processing artifacts. In addition, we introduce BlurRF-Real, a real-world dataset created specifically for evaluation under low-light conditions, where blur frequently occurs.

BlurRF Dataset Image

Visual Comparison

Blurry
DeepDeblurRF-G (ours)

Video Results

BibTeX

@article{choi2025exploiting,
  title={Exploiting Deblurring Networks for Radiance Fields},
  author={Choi, Haeyun and Yang, Heemin and Han, Janghyeok and Cho, Sunghyun},
  journal={CVPR},
  year={2025}
}