RGB-D Surface Reconstruction Using 3D Gaussian Splatting

Recent advancements in RGB-D sensors and deep learning technologies have significantly improved 3D surface reconstruction, which is critical for VR/AR applications. This paper introduces an efficient RGB-D 3D surface reconstruction technique tailored for these applications. The method is based on the assumption that there is a brief period available before content delivery for capturing and reconstructing the environment. It involves a two-step approach: first, using RGB-D SLAM for real-time camera pose estimation, followed by refinement of the reconstruction using 3D Gaussian Splatting. This approach enhances the initial 3D point cloud data and camera pose information to achieve high-quality results. Experimental results demonstrate that this technique substantially improves surface reconstruction quality while enabling rapid processing, allowing users to access detailed and realistic VR/AR content with minimal delay.