Scholarlink

<p>3D Gaussian Splatting (3DGS) has achieved remarkable success in real-time novel view synthesis; however, its reconstruction performance deteriorates when dealing with inputs affected by camera motion or defocus blur. Although recent deblurring methods attempt to address this issue, they predominantly rely on spatial domain features and suffer from the inherent spectral bias of purely spatial optimization, struggling to recover high-frequency texture details. To bridge this gap, we propose FreDeGS, a frequency-guided 3DGS deblurring framework. Our method adopts a physical convolution synthesis baseline to decouple static scene geometry from complex blur. Within this architecture, we introduce the Statistics-Wavelet Gated Multi-scale Adapter (SWGMA), leveraging channel statistics and spatial wavelet energy to enhance high-frequency perception beyond purely spatial constraints. Additionally, an adaptive frequency-aware supervision strategy, driven by the Discrete Wavelet Transform (DWT) and a target-decay mechanism, balances optimization gradients across frequency bands. Extensive experiments on the real-world Deblur-NeRF benchmark demonstrate that FreDeGS improves reconstruction quality under both camera motion and defocus blur. The source code will be made publicly available at <a href="https://github.com/Chiffin-0816/FreDeGS/">https://github.com/Chiffin-0816/FreDeGS/</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FreDeGS: frequency-guided 3D gaussian splatting for scene deblurring

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Xiaofeng Quan,
  • Junzhe Wan,
  • Chao Cai,
  • Qiang Chen,
  • Amin Mao,
  • Yuming Fang

摘要

3D Gaussian Splatting (3DGS) has achieved remarkable success in real-time novel view synthesis; however, its reconstruction performance deteriorates when dealing with inputs affected by camera motion or defocus blur. Although recent deblurring methods attempt to address this issue, they predominantly rely on spatial domain features and suffer from the inherent spectral bias of purely spatial optimization, struggling to recover high-frequency texture details. To bridge this gap, we propose FreDeGS, a frequency-guided 3DGS deblurring framework. Our method adopts a physical convolution synthesis baseline to decouple static scene geometry from complex blur. Within this architecture, we introduce the Statistics-Wavelet Gated Multi-scale Adapter (SWGMA), leveraging channel statistics and spatial wavelet energy to enhance high-frequency perception beyond purely spatial constraints. Additionally, an adaptive frequency-aware supervision strategy, driven by the Discrete Wavelet Transform (DWT) and a target-decay mechanism, balances optimization gradients across frequency bands. Extensive experiments on the real-world Deblur-NeRF benchmark demonstrate that FreDeGS improves reconstruction quality under both camera motion and defocus blur. The source code will be made publicly available at https://github.com/Chiffin-0816/FreDeGS/.