TraGraph-GS: Trajectory Graph-based Gaussian Splatting for Arbitrary Large-Scale Scene Rendering.

Area of Science:

• 3D Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Novel view synthesis for large-scale scenes is challenging due to rigid partitioning and Gaussian overlap issues in existing methods.
• Current techniques struggle with arbitrary camera trajectories and texture distortion when merging scene regions.

Purpose of the Study:

• To develop a generalized method for high-precision novel view synthesis in arbitrarily large-scale 3D scenes.
• To address limitations of spatial partitioning and Gaussian overlap in existing Gaussian splatting methods.

Main Methods:

• Proposed TraGraph-GS, utilizing a trajectory graph for spatial partitioning and scene reconstruction.
• Introduced a graph-based partitioning method with regularization for enhanced texture and distant object rendering.
• Implemented a progressive rendering strategy to mitigate artifacts from Gaussian overlap.

Main Results:

• Demonstrated superior performance and efficiency across aerial, ground, and cross-view datasets.
• Achieved significant PSNR improvements (0.36 dB aerial, 1.24 dB ground, 0.44 dB cross-view) on a single 24GB GPU.
• Outperformed state-of-the-art approaches in novel view synthesis quality.

Conclusions:

• TraGraph-GS effectively enables high-precision rendering for large-scale scenes with arbitrary camera paths.
• The method generalizes well and offers a significant improvement over existing techniques.
• Achieved state-of-the-art results with remarkable efficiency on standard hardware.