A Practical Approach to Indoor Space Reconstruction Using Gaussian Splatting
Jong-Hwan Bae, Sanghun Park · Journal of the Korea Computer Graphics Society · 2025
This study proposes an integrated method to improve the efficiency and accuracy of large-scale indoor space 3D reconstruction based on 3D Gaussian Splatting (3DGS) technology. The process consists of three stages: data preprocessing, object removal, and point cloud merging. First, HyperIQA and ResNet50 models are used to effectively remove low-quality or duplicate images from the dataset, thereby constructing a dataset optimized for 3DGS. Second, Grounded-SAM2 and ProPainter are combined to detect and remove unnecessary objects present in the scene, ensuring visual consistency through inpainting. Finally, to merge the spaces (PLY files) reconstructed in multiple parts, initial alignment is performed using CloudCompare, and the final merging is completed by applying precise transformation matrices with the GaussReg algorithm. This two-step registration approach maintains structural consistency even in spaces with complex shapes and few feature points. This study enhances the practical applicability of 3DGS technology by integrating processes from image preprocessing to object removal and precise registration, and is significant in that it lays the foundation for efficient indoor space data construction in the fields of digital twins and the metaverse.