DRGaussian: Deformable Gaussian Splatting for High-fidelity Dynamic Scene Reconstruction with Depth Regularization

Dynamic scene reconstruction struggles with geometric consistency under sparse viewpoints due to inherent limitations in explicit point-based representations. While Gaussian Splatting excels in rendering efficiency, its point-based paradigm amplifies motion artifacts and structural degradation with limited inputs. We propose DRGaussian, a deformable Gaussian splatting framework for dynamic scene reconstruction. We design effective deformation constraints to smooth the motion field of Gaussian splats. To mitigate degradation in scene geometry, we propose a depth regularization loss to help Gaussians learn structure details of objects based on patch-level and image-level depth supervision. We conducted experiments on multiple public datasets that contain both synthetic and real scenes. Our model outperforms state-of-the-art methods, achieving better results in color and depth reconstruction for dynamic scenes. The project page is https://drgaussian.gitfhub.io.