SV-2DGS: Optimization of sparse view 3D reconstruction based on 2DGS models

In the field of three-dimensional reconstruction, recovering structure from scene images captured from sparse viewpoints has always been a challenging task. The currently popular 3DGS model often encounters issues of blurry rendered scenes and inconsistencies in Gaussian ellipsoids when processing sparse viewpoint data, leading to increased errors. To address these, this study proposes the SV-2DGS model, an optimization method for three-dimensional reconstruction from sparse views based on 2DGS. This model primarily optimizes the point cloud matching in Colmap and the 2DGS reconstruction process. To capture details from blurry views, SV-2DGS introduces a data augmentation framework within Colmap, employing feature extraction and upsampling techniques to generate high-resolution three-dimensional point clouds and camera poses. Furthermore, to enhance the consistency of reconstruction, SV-2DGS employs Gaussian disc reconstruction techniques, which focus on reconstructing those discs with high proximity values by constructing a proximity graph of the Gaussian disk, thereby achieving a smoother reconstruction outcome. Through case analysis, this study demonstrates that SV-2DGS achieves a 0.399 dB improvement in PSNR compared to the pre-optimized 2DGS model, exhibiting outstanding reconstruction accuracy and detail capture in 24-view and high-clarity scenes. Additionally, experiments conducted on self-collected datasets confirm that the SV-2DGS model is suitable for reconstructing low-pixel scene images captured with non-professional equipment in everyday life.