Reconstructing 3D scenes from video has become increasingly important in applications that require high processing speed and vivid imagery, such as augmented reality, interactive simulations, construction monitoring, and immersive presentations. These applications demand solutions that are both time-efficient and capable of producing high-quality visual results, in line with the growing trend of intelligent visual systems. Traditional methods often rely on Structure-from-Motion (SfM) to estimate camera positions and reconstruct sparse geometry, which is then refined using additional techniques for more detailed modeling. Recently, neural radiance field approaches like NeRF have achieved impressive results in novel view synthesis, but at the cost of extremely high computational demands. To address this limitation, this study proposes a creative hybrid pipeline that combines Colmap and OpenSplat. Colmap provides calibrated camera parameters and initial sparse points, taking advantage of its robustness in noisy data or wide-angle scenarios. Meanwhile, OpenSplat represents the scene using learnable anisotropic 3D Gaussians, leveraging the real-time rendering speed and high quality of Gaussian Splatting. Unlike approaches that rely solely on meshes or black-box models like NeRF, the Colmap-OpenSplat system offers a fully differentiable radiance model, enabling direct optimization from input images and efficient rendering. This not only achieves high reconstruction quality but also significantly reduces computational costs, unlocking broader potential for real-world applications.