This paper analyzes the error in 3D Gaussian Splatting (3D-GS) and proposes an optimal projection strategy to reduce projection errors and improve rendering quality. 3D-GS represents scenes using 3D Gaussians, which are projected onto the image plane for rendering. However, this projection process introduces errors due to the local affine approximation used. The authors analyze the error function and find that it is related to the Gaussian mean. They propose an optimal projection strategy that minimizes projection errors by projecting each Gaussian onto a tangent plane rather than the same plane. This approach reduces artifacts and improves rendering quality. The method is simple to implement and can be adapted to various camera models. Experiments show that the proposed method outperforms the original 3D-GS in terms of rendering quality, particularly in scenarios with short focal lengths. The method is also more robust to changes in the field of view and focal length.This paper analyzes the error in 3D Gaussian Splatting (3D-GS) and proposes an optimal projection strategy to reduce projection errors and improve rendering quality. 3D-GS represents scenes using 3D Gaussians, which are projected onto the image plane for rendering. However, this projection process introduces errors due to the local affine approximation used. The authors analyze the error function and find that it is related to the Gaussian mean. They propose an optimal projection strategy that minimizes projection errors by projecting each Gaussian onto a tangent plane rather than the same plane. This approach reduces artifacts and improves rendering quality. The method is simple to implement and can be adapted to various camera models. Experiments show that the proposed method outperforms the original 3D-GS in terms of rendering quality, particularly in scenarios with short focal lengths. The method is also more robust to changes in the field of view and focal length.