科研成果详情

With the introduction of differentiable rendering frameworks, many reconstruction techniques for dynamic scenes have been proposed. Existing methods have poor reconstruction quality for videos recorded in a single view, including problems such as blurred reconstruction of high-frequency details of objects and inaccurate geometry estimation. Inspired by 3D Gaussian Splatting, we propose a framework SV-Gaussians specifically designed for single-view dynamic scene reconstruction and editing. Our method uses a sparse point cloud of the scene as the initialization of the 3D Gaussian and designs a deformation module to model the change of Gaussian properties over time. We propose a deformation regularization to ensure the consistency of adjacent Gaussian motion. We define color constraints based on obstacle masks, help 3D Gaussians remove obstacles and restore the appearance using contextual information of occluded areas, achieving more realistic object removal than image inpainting methods. Extensive experiments show that our method achieves higher reconstruction quality for single-view scenes than SOTA reconstruction methods and guarantees real-time rendering. More demos and code are available at https://sv-gaussians.github.io.