# Denoise-GS: Self-Supervised Denoising for Sparse-View 3D Gaussian Splatting

## TL;DR

This paper introduces Denoise-GS, a new method that improves 3D image quality when input views are sparse and noisy.

## Contribution

Denoise-GS introduces a two-round optimization framework combining self-supervised denoising with 3D Gaussian splatting for sparse-view scenarios.

## Key findings

- Denoise-GS achieves higher PSNR and SSIM scores compared to other methods in sparse-view 3D reconstruction.
- The method performs well in novel view generation with sparse and noisy input images.

## Abstract

Three-dimensional Gaussian splatting has emerged as a mainstream method in the field of new viewpoint synthesis due to its outstanding performance. However, its generation quality typically degrades significantly when input viewpoints are sparse. The introduction of InstantSplat further improved new viewpoint generation in sparse viewpoint scenarios. Nevertheless, these methods produce suboptimal results in sparse viewpoint scenes with noise and no camera prior. To address this issue, we propose Denoise-GS, a two-round optimization framework combining N2V-UNet denoising with InstantSplat rendering. First, Noise2Void performs self-supervised denoising on the input image. Next, pose grouping is conducted based on InstantSplat rendered results. Finally, a second round of refinement is applied to the UNet through a joint loss function. The final denoised result is then re-rendered to achieve a higher-quality output image. To simulate a real noisy environment, we added Gaussian noise to the input images. Tests on multiple datasets show that, compared with other mainstream methods, our approach produces images with higher PSNR and SSIM. The method performs well in novel view generation when the input images are sparse and noisy, providing an innovative and practical solution for three-dimensional reconstruction.