PGSR: Planar-Based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Deep Learning

Background:

• 3D Gaussian Splatting (3DGS) offers fast rendering but struggles with geometric accuracy due to unstructured point clouds.
• Existing 3DGS surface reconstruction methods yield unsatisfactory mesh quality.
• Ensuring multi-view consistency and geometric precision remains a challenge for 3DGS.

Purpose of the Study:

• To propose a novel planar-based Gaussian splatting reconstruction representation (PGSR) for high-fidelity surface reconstruction and rendering.
• To enhance geometric accuracy and multi-view consistency in 3D Gaussian Splatting.
• To develop a method that balances fast training/rendering with high-quality geometric reconstruction.

Main Methods:

• Introduced an unbiased depth rendering method using Gaussian plane distances and normal maps.
• Implemented single-view geometric, multi-view photometric, and geometric regularization for global accuracy.
• Developed a camera exposure compensation model for varying illumination conditions.

Main Results:

• PGSR achieves high-fidelity rendering and geometric reconstruction.
• The method demonstrates fast training and rendering speeds.
• Experimental results show superior performance compared to 3DGS-based and NeRF-based approaches.

Conclusions:

• PGSR effectively addresses the limitations of existing 3DGS methods for surface reconstruction.
• The proposed approach offers a robust solution for high-quality 3D scene representation.
• PGSR provides a promising direction for future research in real-time 3D reconstruction and rendering.