The paper introduces FreGS, an innovative 3D Gaussian splatting framework designed to address the over-reconstruction issue in novel view synthesis. This issue is tackled through frequency regularization in the frequency space, which is a novel approach to mitigating the deficiencies in 3D Gaussian splatting. FreGS introduces a frequency annealing technique for progressive frequency regularization, which involves a coarse-to-fine Gaussian densification process by annealing the regularization progressively from low-frequency signals to high-frequency signals. This approach is based on the rationale that low-frequency and high-frequency signals encode large-scale and small-scale features, respectively. By minimizing the discrepancy of frequency spectra of the rendered image and the corresponding ground truth, FreGS effectively mitigates the over-reconstruction issue and achieves superior Gaussian densification and novel view synthesis.

In terms of implementation, the paper details the training process of FreGS, which involves using the pixel-level L1 loss in the spatial space along with the D-SSIM term throughout the training process. The optimization process starts with an image resolution that is four times smaller than the original images, and after 500 iterations, the resolution is increased to the original size by upsampling. The training of FreGS is carried out using the Adam optimizer and the PyTorch framework for implementation. Additionally, for rasterization, the custom CUDA kernels used in 3D-GS are retained. The paper also discusses the comparisons with the state-of-the-art methods, where FreGS consistently outperforms 3D-GS in terms of image rendering quality across various real scenes. This superior performance is largely attributed to the proposed progressive frequency regularization, which effectively alleviates the over-reconstruction issue of Gaussians and improves Gaussian densification.