DynSUP: Dynamic Gaussian Splatting From an Unposed Image Pair.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• 3D Gaussian Splatting excels in static scenes with multiple images and known camera poses.
• Dynamic scenes, sparse views, and unknown poses present significant challenges due to limited geometric constraints.

Purpose of the Study:

• To develop a method for dynamic 3D scene reconstruction from only two unposed images.
• To enable high-fidelity novel view synthesis and accurate motion modeling in dynamic environments.

Main Methods:

• Object-level two-view bundle adjustment to decompose dynamic scenes and estimate camera/object motion.
• SE(3) field-driven Gaussian training for fine-grained, per-Gaussian motion modeling.

Main Results:

• Achieved high-fidelity novel view synthesis for dynamic scenes.
• Accurately preserved temporal consistency and object motion.
• Outperformed state-of-the-art methods on synthetic and real-world datasets.

Conclusions:

• The proposed method effectively reconstructs dynamic 3D scenes from minimal, unposed input.
• It overcomes limitations of existing methods in handling dynamic environments and unknown camera poses.