Enhanced 3D Gaussian Splatting for Real-Scene Reconstruction via Depth Priors, Adaptive Densification, and Denoising

Area of Science:

• Computer Vision
• 3D Reconstruction
• Photogrammetry

Background:

• Photorealistic 3D reconstruction is vital for smart cities and cultural heritage.
• Existing methods struggle with accuracy, efficiency, and robustness in complex scenes (e.g., reflective surfaces, vegetation).

Purpose of the Study:

• To propose an enhanced 3D Gaussian Splatting (3DGS) framework.
• To improve reconstruction accuracy, computational efficiency, and robustness in complex 3D scenes.

Main Methods:

• Integrated a depth-aware regularization module using Depth-Anything V2 for geometrically informed optimization.
• Implemented a gradient-driven adaptive densification mechanism for efficient Gaussian adjustments.
• Developed a neighborhood density-based method for detecting and filtering floating artifacts.

Main Results:

• Achieved state-of-the-art performance with PSNR of 34.15 dB and SSIM of 0.9382 on medium scenes.
• Demonstrated real-time rendering speeds over 170 FPS at 1600x900 resolution.
• Showcased superior generalization on challenging materials (water, foliage) and reduced overfitting.

Conclusions:

• The enhanced 3DGS framework significantly improves 3D reconstruction quality and efficiency.
• Depth regularization and gradient-sensitive adaptation are critical for performance gains.
• Optimal input resolution scaling (1/4-1/2) balances fidelity and efficiency, though large-scale memory consumption requires further research.