Hie4DGS: Hierarchical 4D Gaussian Splatting From Monocular Dynamic Video.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Computer Graphics

Background:

• Monocular dynamic video reconstruction is challenging due to limited observations and complex 3D motions.
• Existing dynamic 3D Gaussian splatting methods often fail in monocular settings, relying on multi-camera cues or neglecting structural coherence.

Purpose of the Study:

• To propose Hie4DGS, a novel hierarchical structure representation for modeling complex dynamic motions from monocular videos.
• To address the limitations of current dynamic 3D Gaussian splatting techniques in monocular scenarios.

Main Methods:

• Decomposing scene motion into multiple granularities using a hierarchical structure.
• Leveraging hierarchical semantic segmentation to group 3D Gaussians and initialize motion with depth and tracking priors.
• Introducing a structure rendering loss for enhanced motion structure consistency.

Main Results:

• Achieved significant improvements in rendering quality on monocular video datasets with complex motions.
• Demonstrated superior performance compared to state-of-the-art dynamic Gaussian methods in monocular reconstruction.
• Successfully reduced motion ambiguity through the proposed hierarchical representation and rendering loss.

Conclusions:

• Hie4DGS effectively models complex dynamic motions in monocular videos using a hierarchical structure.
• The proposed method enhances the rendering quality and structural coherence of dynamic 3D Gaussian splatting.
• This work offers a promising solution for challenging monocular dynamic video reconstruction tasks.