GS-MSDR: Gaussian Splatting with Multi-Scale Deblurring and Resolution Enhancement

Area of Science:

• Computer Vision
• Computer Graphics
• Image Processing

Background:

• 3D Gaussian Splatting (3DGS) excels at scene reconstruction and novel view synthesis on clean datasets.
• Real-world images suffer degradations (camera shake, motion blur, defocus) hindering 3D reconstruction accuracy, especially fine details.
• Existing deblurring methods often fail with multiple blur types and resolution degradation.

Purpose of the Study:

• To develop a robust framework for high-fidelity 3D reconstruction from severely degraded real-world images.
• To enhance 3D Gaussian Splatting (3DGS) performance by integrating advanced deblurring and resolution enhancement techniques.
• To address limitations of single-blur deblurring methods in complex, multi-degradation scenarios.

Main Methods:

• Proposed Gaussian Splatting with Multi-Scale Deblurring and Resolution Enhancement (GS-MSDR) framework.
• Developed a Multi-scale Adaptive Attention Network (MAAN) for fusing multi-scale features.
• Incorporated Multi-modal Context Adapter (MCA) and adaptive spatial pooling for refined feature representation.
• Utilized Hierarchical Progressive Kernel Optimization (HPKO) for layer-wise optimization and precise detail reconstruction.

Main Results:

• GS-MSDR significantly outperforms state-of-the-art methods across diverse degraded image scenarios.
• Achieved superior deblurring quality and highly accurate 3D reconstruction, preserving fine details.
• Demonstrated efficient rendering capabilities within the 3DGS framework.

Conclusions:

• GS-MSDR provides a robust solution for 3D reconstruction from challenging, degraded real-world images.
• The integrated multi-scale deblurring and resolution enhancement effectively recovers fine details lost in complex degradations.
• This framework advances the applicability of 3D Gaussian Splatting in practical, uncurated environments.