Skip to content
Book Open access

LDSG: Learning Decoupled Structured Gaussians for Improved Dynamic Rendering and Reconstruction

Nov 2026 · 0 citations · 46 references

Abstract

Learning dynamic object representations from monocular videos is a crucial challenge in computer vision and graphics, demanding accurate reconstruction and realistic rendering. The recent 3D Gaussian Splatting method offers a powerful representation for dynamic rendering. However, the discrete and unstructured nature of Gaussians makes accurate surface reconstruction difficult. Some methods use squashed 3D Gaussians to improve geometric accuracy, sacrificing rendering fidelity, while others adopt dual-model architectures with high computational cost. To address these limitations, we propose LDSG, which learns Decoupled Structured Gaussians to jointly improve rendering and reconstruction. Specifically, we introduce a dual canonical Gaussian representation to decouple the conflict between appearance and geometry, where the covariance of the geometric Gaussians is derived by inheriting structural priors from the rendering Gaussians. We then propose an Anchor-GS guided hierarchical deformation to mitigate interference between rendering and reconstruction in single-level deformation fields. Subsequently, a spatiotemporally-aware anchor densification scheme adaptively enhances model detail representation. Moreover, we design a virtual-view depth consistency regularization to improve geometric accuracy, which enforces cross-view geometric coherence using synthesized virtual viewpoints. These components form a unified and efficient framework that avoids maintaining separate rendering and reconstruction models while preserving their complementary strengths. Extensive experiments demonstrate that our method achieves state-of-the-art performance in dynamic rendering and reconstruction while maintaining high efficiency.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.