Pano-GS: sparse-view 3D Gaussian splatting for scene reconstruction with panoramic depth priors
Abstract
Sparse-view 3D Gaussian Splatting (3DGS) often suffers from unstable initialization and limited geometric continuity, particularly in open scenes with sparse viewpoint coverage. To address these issues, we propose Pano-GS, a sparse-view 3D reconstruction method based on panoramic depth priors. The proposed method leverages the full 360° field of view of panoramic images of panoramic images by partitioning each panorama into virtual perspective views and introducing camera rig constraints for pose initialization. Initial camera poses are estimated with VGGT, while panoramic depth prediction serves as a geometric prior to generate a consistent initial point cloud through multi-view depth scale alignment. Based on this initialization, multi-view depth back-projection is further incorporated for Gaussian initialization and depth supervision. In addition, a 2D Gaussian representation is adopted to improve surface optimization and geometric continuity. Experimental results on Replica, 360Roam, and SZCity datasets demonstrate that, with only a small number of panoramic images as input, the proposed method effectively improves reconstruction stability and geometric quality under sparse-view settings.