Recently, 3D Gaussian splatting (3DGS) has gained considerable attentions in the field of novel view synthesis due to its fast performance while yielding the excellent image quality. However, 3DGS in sparse-view settings (e.g., three-view inputs) often faces with the problem of overfitting to training views, which significantly drops the visual quality of novel view images. Many existing approaches have tackled this issue by using strong priors, such as 2D generative contextual information and external depth signals. In contrast, this paper introduces a prior-free method, so-called DropGaussian, with simple changes in 3D Gaussian splatting. Specifically, we randomly remove Gaussians during the training process in a similar way of dropout, which allows non-excluded Gaussians to have larger gradients while improving their visibility. This makes the remaining Gaussians to contribute more to the optimization process for rendering with sparse input views. Such simple operation effectively alleviates the overfitting problem and enhances the quality of novel view synthesis. By simply applying DropGaussian to the original 3DGS framework, we can achieve the competitive performance with existing prior-based 3DGS methods in sparse-view settings of benchmark datasets without any additional complexity. The code and model are publicly available at: https://github.com/DCVL-3D/DropGaussian_release.

In recent years, 3D Gaussian splatting (3DGS) has shown high-fidelity rendering results in real-time. However, 3DGS often encounters the overfitting problem under sparse-view conditions due to insufficient cross-view constraints. In this letter, to mitigate this limitation, we focus on the effect of Gaussian shapes on the scene reconstruction from sparse input views. The key idea is to allow each Gaussian to adaptively select its shape in accordance with the scene structure. Specifically, we propose to put a learnable parameter into Gaussian attributes, which indicates the probability of each shape. This indicator is optimized with other attributes while making each Gaussian change its shape to 1D, 2D, and 3D for representing edges, planar surfaces, and volumetric regions, respectively. Based on a geometrically accurate representation, the proposed method consequently alleviates the model from overfitting to a limited set of training views. Furthermore, we apply a depth regularization scheme within a set of selected pixels to precisely constrain positions of Gaussians. Experimental results on benchmark datasets show that the proposed method effectively improves the performance of novel view synthesis under sparse input views. The code and model are publicly available at: https://github.com/DCVL3D/SSS-release.