DispFlow-GS: Displacement Flow Supervision with Motion Disentangling for Monocular Deformable 3D Gaussian Splatting

TL;DR

The paper introduces DispFlow-GS, a method that improves motion modeling in monocular deformable 3D Gaussian Splatting by using displacement flow supervision instead of traditional Gaussian flow alignment. It disentangles camera motion from scene motion through intermediate-view rendering and proposes a new motion-aware evaluation metric, Deformation-Rendering Consistency (DRC), to better assess motion fidelity. Experiments show significant gains in motion accuracy and consistency, even when image-based metrics remain largely unchanged.

Why It Matters

Accurate motion modeling is critical for robotics and autonomous systems operating in dynamic environments. This work addresses a key gap in evaluating motion quality by introducing a metric that aligns with motion fidelity rather than image appearance, enabling more reliable assessment of dynamic scene reconstruction. The approach improves the robustness and interpretability of 3D reconstruction models in real-world applications.

Key Findings

• Gaussian-flow-based motion supervision provides limited gains due to a domain mismatch between rendered flow and optical flow.
• Displacement flow supervision offers more stable and direct optimization signals by projecting 3D Gaussian displacements onto the image plane.
• Deformation-Rendering Consistency (DRC) effectively measures motion-aware improvement, revealing a disconnect between motion accuracy and standard image-based metrics.