FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization

FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization

8 Apr 2024 | Jiahui Zhang, Fangneng Zhan, Muyu Xu, Shijian Lu, Eric Xing
The paper introduces FreGS, a novel 3D Gaussian splatting technique that addresses the over-reconstruction issue in 3D Gaussian splatting (3D-GS) by incorporating progressive frequency regularization. Over-reconstruction occurs when high-variance image regions are covered by a few large Gaussians, leading to blur and artifacts in the rendered images. FreGS uses a frequency annealing technique to perform coarse-to-fine Gaussian densification, gradually leveraging low-to-high frequency components in the Fourier space. This approach minimizes the discrepancy between the frequency spectrum of the rendered image and the ground truth, improving the quality of Gaussian densification and novel view synthesis. Experiments on multiple benchmarks, including Mip-NeRF360, Tanks-and-Temples, and Deep Blending, demonstrate that FreGS consistently outperforms state-of-the-art methods in terms of image quality and novel view synthesis. The proposed FreGS technique provides a novel perspective on addressing over-reconstruction in 3D Gaussian splatting, offering superior performance and visual quality.The paper introduces FreGS, a novel 3D Gaussian splatting technique that addresses the over-reconstruction issue in 3D Gaussian splatting (3D-GS) by incorporating progressive frequency regularization. Over-reconstruction occurs when high-variance image regions are covered by a few large Gaussians, leading to blur and artifacts in the rendered images. FreGS uses a frequency annealing technique to perform coarse-to-fine Gaussian densification, gradually leveraging low-to-high frequency components in the Fourier space. This approach minimizes the discrepancy between the frequency spectrum of the rendered image and the ground truth, improving the quality of Gaussian densification and novel view synthesis. Experiments on multiple benchmarks, including Mip-NeRF360, Tanks-and-Temples, and Deep Blending, demonstrate that FreGS consistently outperforms state-of-the-art methods in terms of image quality and novel view synthesis. The proposed FreGS technique provides a novel perspective on addressing over-reconstruction in 3D Gaussian splatting, offering superior performance and visual quality.
Reach us at info@study.space