Graph-Based 3DGS-to-Mesh Reconstruction

Constructed KNN graphs over 3D Gaussian positions with hybrid edge weights combining spatial distance, opacity, scale, and covariance features.

Developed covariance-guided iterative surface score propagation to identify surface Gaussians and filter floaters, enabling cleaner point cloud extraction.

Applied Poisson Surface Reconstruction on cleaned point clouds; surveyed SuGaR, 2DGS, GOF, and L-GGSC for reconstruction quality analysis.

Method addresses a key limitation of vanilla 3DGS — lack of explicit surface representation — relevant to both NVS quality and downstream 3D applications.