Frequency-Aware Uncertainty Gaussian Splatting for Dynamic Scene Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• 3D Gaussian splatting shows promise for dynamic scene reconstruction.
• Existing methods struggle with complex scenes due to rigid deformation models and insufficient high-frequency motion learning, causing artifacts.

Purpose of the Study:

• To propose a novel method, Frequency-aware Uncertainty Gaussian Splatting (FUGS), for adaptive dynamic scene reconstruction.
• To address limitations in modeling motion correlation and learning high-frequency details in dynamic scenes.

Main Methods:

• Developed an Uncertainty-aware Deformation Model (UDM) to capture motion correlations using learnable uncertainty relations between Gaussian points.
• Introduced Dynamic Spectrum Regularization (DSR) for coarse-to-fine Gaussian densification via frequency filtering, adapting attributes based on scene complexity.

Main Results:

• FUGS achieves high-fidelity reconstruction of complex dynamic scenes.
• The method demonstrates real-time rendering capabilities.
• Experiments show significant superiority over state-of-the-art methods on synthetic and real-world datasets.

Conclusions:

• FUGS offers a more adaptable and robust approach to dynamic scene reconstruction.
• The proposed methods effectively handle motion uncertainties and improve high-frequency detail learning.
• FUGS represents a significant advancement in reconstructing complex dynamic environments.