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

Area of Science:

• Computer Vision
• Computer Graphics
• Image Processing

Background:

• Three-dimensional Gaussian splatting is a leading technique for novel viewpoint synthesis.
• Existing methods like InstantSplat struggle with sparse viewpoints, noise, and lack of camera prior.
• Degradation in generation quality is a significant issue in sparse input scenarios.

Purpose of the Study:

• To develop a robust framework for high-quality novel view synthesis from sparse and noisy input data.
• To overcome the limitations of current 3D Gaussian splatting methods in challenging real-world conditions.
• To improve the accuracy and visual fidelity of 3D reconstruction.

Main Methods:

• A two-round optimization framework, Denoise-GS, is proposed.
• Incorporates Noise2Void (N2V) for self-supervised denoising of input images.
• Combines N2V-UNet denoising with InstantSplat rendering, including pose grouping and joint loss refinement.

Main Results:

• Denoise-GS significantly improves image quality, achieving higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).
• The method demonstrates superior performance in novel view generation with sparse and noisy input images compared to mainstream approaches.
• Successfully simulated noisy environments by adding Gaussian noise to input images for testing.

Conclusions:

• Denoise-GS offers an innovative and practical solution for 3D reconstruction and novel view synthesis.
• The framework effectively addresses the challenges posed by sparse and noisy input data.
• Achieves state-of-the-art results, enhancing the applicability of 3D Gaussian splatting in real-world scenarios.