The paper "CoR-GS: Sparse-View 3D Gaussian Splatting via Co-Regularization" addresses the issue of overfitting in sparse-view 3D Gaussian Splatting (3DGS) and proposes a novel co-regularization approach to improve its performance. The authors observe that when training two 3D Gaussian radiance fields with sparse views, there is a significant increase in point disagreement and rendering disagreement during the densification process, which can lead to inaccurate reconstruction. They quantify these disagreements and demonstrate their negative correlation with accurate reconstruction quality. Based on this observation, they propose CoR-GS, which includes two main components: co-pruning and pseudo-view co-regularization. Co-pruning identifies and prunes Gaussians that exhibit high point disagreement in inaccurate positions, while pseudo-view co-regularization suppresses rendering disagreement by sampling pseudo views and comparing rendered pixels. The method is evaluated on multiple datasets (LLFF, Mip-NeRF360, DTU, and Blender) and shows significant improvements in scene geometry reconstruction and novel view synthesis quality compared to state-of-the-art methods. The paper also provides a detailed analysis of the effectiveness of each component and discusses future work directions.The paper "CoR-GS: Sparse-View 3D Gaussian Splatting via Co-Regularization" addresses the issue of overfitting in sparse-view 3D Gaussian Splatting (3DGS) and proposes a novel co-regularization approach to improve its performance. The authors observe that when training two 3D Gaussian radiance fields with sparse views, there is a significant increase in point disagreement and rendering disagreement during the densification process, which can lead to inaccurate reconstruction. They quantify these disagreements and demonstrate their negative correlation with accurate reconstruction quality. Based on this observation, they propose CoR-GS, which includes two main components: co-pruning and pseudo-view co-regularization. Co-pruning identifies and prunes Gaussians that exhibit high point disagreement in inaccurate positions, while pseudo-view co-regularization suppresses rendering disagreement by sampling pseudo views and comparing rendered pixels. The method is evaluated on multiple datasets (LLFF, Mip-NeRF360, DTU, and Blender) and shows significant improvements in scene geometry reconstruction and novel view synthesis quality compared to state-of-the-art methods. The paper also provides a detailed analysis of the effectiveness of each component and discusses future work directions.