Consistent 4D Appearance Editing with Gaussian Splatting

Area of Science:

• Computer Vision
• Computer Graphics
• Scientific Visualization

Background:

• Dynamic scene editing with 4D Gaussian Splatting (4DGS) faces challenges with spatiotemporal inconsistencies, known as Gaussian drifting.
• These inconsistencies degrade the quality of edits and temporal coherence in dynamic scenes.

Purpose of the Study:

• To develop a comprehensive framework to address spatiotemporal inconsistencies in 4D Gaussian Splatting editing.
• To improve the quality and temporal coherence of edited dynamic scenes.

Main Methods:

• Introduced a prior-guided, multi-stage reconstruction pipeline fusing geometric and motion priors for a stable scene foundation.
• Applied a universal trajectory-preserving technique to decouple appearance optimization from deformation, safeguarding motion quality.

Main Results:

• The proposed framework systematically tackles both reconstruction-induced and editing-induced inconsistencies.
• Achieved state-of-the-art, temporally consistent editing on diverse dynamic scenes.

Conclusions:

• The method effectively resolves Gaussian drifting in 4DGS editing by addressing foundational reconstruction and editing process errors.
• Demonstrates superior performance over monolithic approaches for dynamic scene editing.