SFR-GS: Spatial-Frequency Domain Regularization for 3D Gaussian Splatting

Guanhua Wu, Jiahao Wu, Kaiqiang Xiong, Zhanke Wang, Ronggang Wang

Abstract: 3D Gaussian Splatting(3DGS) has demonstrated impressive performance in real-time novel view synthesis. However, it frequently suffers over-reconstruction issue in intricate scenes, where high-frequency details are only covered by a few large Gaussians, leading to blurring and artifacts. To address this issue, we propose a plug-and-play method named SFR, which introduces regularization for the high-frequency details in both the spatial and frequency domains of the image. Specifically, in the spatial domain, we propose a gradient-guided densification strategy that adaptively splits Gaussians along image structures to better align with object boundaries and textures. Additionally, we design an adaptive loss weighting scheme that leverages image gradients to emphasize structurally informative regions. In the frequency domain, we introduce a novel frequency stability measure to guide adaptive spectral regularization, enabling the model to focus on unstable high-frequency components while preserving low-frequency consistency. Extensive qualitative and quantitative experiments on diverse baselines and challenging datasets such as Mip-NeRF360, Tanks&Temples, and DeepBlending, demonstrate the effectiveness and generalizability of our method to enhance the reconstruction quality of high-frequency details of current 3DGS methods.

External IDs:dblp:conf/prcv/WuWXWW25