Real-time Gaussian Splatting for Dynamic Reconstruction in Stationary Monocular Cameras

With the development of differentiable rendering techniques such as Gaussian Splatting, the quality of 3D reconstruction has gradually improved. Nowadays on social platforms, the number of videos shot with fixed cameras increases significantly. However, current 3D reconstruction methods require multi-view observation as input, and perform poorly on single-view videos, such as reconstruction blur for moving objects. In order to solve the problem of dynamic scene reconstruction under a single perspective, we propose a network based on dynamic Gaussian Splatting and implement the task of removing objects during reconstruction process. Our method uses the depth projection to obtain a point cloud as the initialization of Gaussians, and designs a lightweight Gaussian deformation module to obtain the changes of each Gaussian over time. The deformation module utilizes discrete feature planes as encoder, which significantly reduces training time. In order to remove dynamic objects, we use an object mask and design a loss function to restore the color of occluded area. We conducted extensive experiments to evaluate our method and SOTA reconstruction methods. Our method not only significantly improves the reconstruction quality and training time, but also achieves real-time rendering.