Panoramic Image Driven Point Cloud Initialization for 3D Reconstruction

The ability to reconstruct immersive and realistic three-dimensional scenes plays a fundamental role in advancing virtual reality, digital twins, and related fields. With the rapid development of differentiable rendering frameworks, the reconstruction quality of static scenes has been significantly improved. However, we observe that the challenge of insufficient initialization has been largely overlooked in existing studies, while at the same time heavily relying on dense multi-view imagery that is difficult to obtain. To address these challenges, we propose a pipeline for text driven 3D scene generation, which employs panoramic images as an intermediate representation and integrates with 3D Gaussian Splatting to enhance reconstruction quality and efficiency. Our method introduces an improved point cloud initialization using Fibonacci lattice sampling of panoramic images, combined with a dense perspective pseudo label strategy for teacher–student distillation supervision, enabling more accurate scene geometry and robust feature learning without requiring explicit multi-view ground truth. Extensive experiments validate the effectiveness of our method, consistently outperforming state-of-the-art methods across standard reconstruction metrics.