Abstract
3D Gaussian Splatting has garnered extensive attention and application inreal-time neural rendering. Concurrently, concerns have been raised about thelimitations of this technology in aspects such as point cloud storage,performance , and robustness in sparse viewpoints , leading to variousimprovements. However, there has been a notable lack of attention to theprojection errors introduced by the local affine approximation inherent in thesplatting itself, and the consequential impact of these errors on the qualityof photo-realistic rendering. This paper addresses the projection errorfunction of 3D Gaussian Splatting, commencing with the residual error from thefirst-order Taylor expansion of the projection function $\phi$. The analysisestablishes a correlation between the error and the Gaussian mean position.Subsequently, leveraging function optimization theory, this paper analyzes thefunction's minima to provide an optimal projection strategy for GaussianSplatting referred to Optimal Gaussian Splatting. Experimental validationfurther confirms that this projection methodology reduces artifacts, resultingin a more convincingly realistic rendering.