Advances in Dynamic Scene Reconstruction

The field of dynamic scene reconstruction is moving towards more sophisticated methods for modeling and restoring complex scenes. Researchers are focusing on developing techniques that can accurately capture motion, structure, and appearance from limited observations. A key direction is the use of physical models and geometric priors to improve the quality of restored videos and reconstructed scenes. Another important trend is the development of methods that can explicitly model heterogeneous motion patterns and decouple static and dynamic elements in scene representations. These innovations are leading to significant improvements in reconstruction fidelity, temporal consistency, and motion separation. Noteworthy papers include: PMR, which proposes a physical model-driven multi-stage restoration framework for turbulent dynamic videos. GeoMoE, which introduces a mixture-of-experts approach for motion field modeling in two-view geometry. SplitGaussian, which presents a novel framework for reconstructing dynamic scenes via visual geometry decomposition. Laplacian Analysis Meets Dynamics Modelling, which proposes a dynamic 3D Gaussian Splatting framework with hybrid explicit-implicit functions.

Sources

PMR: Physical Model-Driven Multi-Stage Restoration of Turbulent Dynamic Videos

GeoMoE: Divide-and-Conquer Motion Field Modeling with Mixture-of-Experts for Two-View Geometry

SplitGaussian: Reconstructing Dynamic Scenes via Visual Geometry Decomposition

Laplacian Analysis Meets Dynamics Modelling: Gaussian Splatting for 4D Reconstruction

Built with on top of