This paper introduces Trim 3D Gaussian Splatting (TrimGS), a novel method for reconstructing accurate 3D geometry from images. Unlike previous approaches that focus on strong geometry regularization, TrimGS proposes a Gaussian trimming strategy to selectively remove inaccurate or redundant Gaussians while preserving accurate structures. The method evaluates the contribution of individual 3D Gaussians using a metric inspired by alpha-blending, which helps in identifying and removing Gaussians with low contributions. Additionally, TrimGS maintains relatively small Gaussian scales to better represent intricate details and optimize high-frequency regions. The method is compatible with existing geometry regularization strategies and can be integrated with both 3DGS and 2DGS. Experimental results on the DTU dataset demonstrate that TrimGS consistently yields more accurate geometry and higher perceptual quality compared to original 3DGS and state-of-the-art 2DGS methods. The paper also includes a theoretical analysis showing that large Gaussians have limited capacity to represent detailed geometry and high-frequency regions, and provides insights into the effectiveness of the proposed contributions and scale control strategies.This paper introduces Trim 3D Gaussian Splatting (TrimGS), a novel method for reconstructing accurate 3D geometry from images. Unlike previous approaches that focus on strong geometry regularization, TrimGS proposes a Gaussian trimming strategy to selectively remove inaccurate or redundant Gaussians while preserving accurate structures. The method evaluates the contribution of individual 3D Gaussians using a metric inspired by alpha-blending, which helps in identifying and removing Gaussians with low contributions. Additionally, TrimGS maintains relatively small Gaussian scales to better represent intricate details and optimize high-frequency regions. The method is compatible with existing geometry regularization strategies and can be integrated with both 3DGS and 2DGS. Experimental results on the DTU dataset demonstrate that TrimGS consistently yields more accurate geometry and higher perceptual quality compared to original 3DGS and state-of-the-art 2DGS methods. The paper also includes a theoretical analysis showing that large Gaussians have limited capacity to represent detailed geometry and high-frequency regions, and provides insights into the effectiveness of the proposed contributions and scale control strategies.