3DGS-HD: Elimination of Unrealistic Artifacts in 3D Gaussian Splatting

3D Gaussian Splatting (3DGS) can achieve higher-quality novel view synthesis (NVS) results in a relatively short period of time. Nevertheless, when sampling rates are not aligned with the camera track, unreal artifacts that are not part of the view often emerge, including floaters. This phenomenon has a significant impact on the quality of the reconstruction results. Furthermore, existing regularization methods are ineffective in terms of both inference speed and optimization. Accordingly, this study proposes the elimination of unrealistic artifacts in 3D Gaussian Splatting (3DGS-HD), a scene optimization method based on the 3D diffusion model. This method optimizes the adaptive density mechanism by deeply understanding the local structural characteristics of the reconstructed objects, removes unrealistic artifacts, and significantly improves the reconstruction quality. Furthermore, it ensures consistency between 2D and 3D views. The results of the experiments demonstrate that this method produces notable improvements in optimization, particularly in the context of object-level reconstruction tasks.