SMDGS: Scale-Aligned Monocular Depth-Guided 3D Gaussian Splatting for Rendering and Surface Reconstruction

Area of Science:

• Computer Vision
• 3D Reconstruction
• Computer Graphics

Background:

• 3D Gaussian Splatting (3DGS) shows potential for surface reconstruction.
• Existing 3DGS methods struggle with accuracy and novel view synthesis (NVS) due to unstructured point clouds.
• Monocular depth estimation offers geometric cues but suffers from scale ambiguity.

Purpose of the Study:

• To develop a novel framework for high-quality 3D surface reconstruction and NVS.
• To address the limitations of unstructured point clouds in 3DGS.
• To leverage monocular depth estimation for improved geometric accuracy.

Main Methods:

• Proposed a scale-aligned monocular depth-guided 3DGS framework.
• Implemented a K-Nearest Neighbor (KNN)-based depth alignment using Structure from Motion (SfM) point clouds for regularization.
• Introduced a pseudo-mesh-based multi-view consistency module for surface refinement.
• Utilized a pixel-level isotropic gradient aware method to guide Gaussian growth.

Main Results:

• Achieved accurate surface reconstruction across diverse datasets (indoor, outdoor, object-centered).
• Demonstrated excellent Novel View Synthesis (NVS) performance.
• Significantly improved rendering quality and geometric representation compared to existing methods.

Conclusions:

• The proposed scale-aligned monocular depth-guided 3DGS framework effectively enhances surface reconstruction and NVS quality.
• Combining geometric prior regularization and consistency supervision is key to overcoming 3DGS limitations.
• The method shows broad applicability for various scene types.