ResearchPod Summary
Extending 3D Gaussian Splatting (3DGS) to dynamic scenes often involves decomposing the scene into static background and dynamic foreground components. However, existing deformation networks frequently fail to accurately distinguish between these two, leading to pseudo-static residuals—where moving objects are incorrectly modeled as static—which degrades both background purity and foreground motion fidelity. This paper asks how explicit motion awareness can be injected into these deformation networks to improve this decomposition.
The authors propose MVFusion-GS, a framework that enhances standard deformation networks with two complementary, motion-aware mechanisms that operate in the feature space without requiring changes to the underlying deformation heads. First, the Motion-Variance Guided Refinement (MVG) module computes per-Gaussian motion statistics (position, rotation, and scale variance) across time to create a global trajectory signature and a local motion intensity dictionary. These act as explicit priors for dynamic-static separation. Second, the MotionFormer Temporal Attention (MFTA) module uses a Transformer-based cross-attention mechanism to aggregate motion cues from neighboring timesteps, allowing the network to model local temporal dependencies and improve deformation consistency.
MVFusion-GS achieves state-of-the-art performance on both dynamic scene reconstruction and distractor-free reconstruction benchmarks. By explicitly modeling motion intensity and temporal context, the method successfully re-attributes pseudo-static Gaussians to the dynamic branch, resulting in cleaner static backgrounds and more accurate foreground motion capture. The plug-in nature of these modules allows them to be integrated into existing deformation-based pipelines, consistently improving rendering quality and temporal stability across various datasets like Neu3D and NeRF On-the-go.
This work addresses a fundamental limitation in current dynamic 3DGS methods: the lack of explicit motion awareness in deformation fields. By providing a lightweight, feature-space refinement that leverages both long-term trajectory statistics and short-term temporal context, the authors offer a robust way to improve scene decomposition. This is particularly valuable for applications requiring high-fidelity reconstruction from casual, unconstrained video captures where transient distractors often contaminate the static background.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.