We tackle the challenge of reconstructing photorealistic 3D scenes from sparse, unposed images. MP-GS uses the pretrained DUSt3R model for dense geometry initialization, enabling geometrically aligned Gaussians and improved pose refinement via dense matching and photometric loss. GBA-GS extends Gaussian Splatting into a scalable, pose-free Global SfM framework, jointly optimizing camera poses and Gaussians from random initialization without training. Experiments on DTU, LLFF, MipNeRF360, Replica, and Tanks-and-Temples show superior pose recovery and geometric consistency compared to prior pose-free 3DGS methods. Our work bridges SfM with photorealistic Gaussian Splatting, integrating correspondence and photometric cues for scalable 3D scene reconstruction.
