Temporally Aware Densification for Dynamic 3D Gaussian Splatting

Existing dynamic 3D Gaussian Splatting (3DGS) methods typically inherit densification strategies designed for static scenes. This creates a fundamental mismatch: dynamic Gaussians often have short temporal lifespans, meaning they are only visible for a few frames. Because standard densification relies on accumulating positional gradients over a fixed interval, these short-lived Gaussians receive insufficient supervision, leading to under-reconstructed, blurry dynamic regions.

The authors propose a unified framework to make densification temporally aware through three primary components:

1. Visibility-Aware Densification (VAD): Instead of simple gradient accumulation, VAD weights positional gradients by each Gaussian's per-frame opacity. This ensures that the densification signal is derived primarily from frames where the Gaussian is actually visible.
2. Temporally-Adaptive Thresholding (TAT): Recognizing that short-lived Gaussians struggle to reach standard gradient thresholds, TAT dynamically lowers the densification threshold for Gaussians with shorter lifespans, promoting their refinement.
3. Temporal Offset Warping (TOW): To handle highly dynamic motion, TOW non-linearly warps the input time coordinate around a Gaussian's temporal center. By allocating more deformation capacity near this center, the model can better capture complex, rapid motion, effectively extending the lifespan of these Gaussians and allowing for more robust densification.

The proposed framework significantly improves the visual quality of dynamic regions across multiple benchmark datasets. The VAD module acts as a plug-and-play component, consistently enhancing the performance of various existing dynamic 3DGS methods. By explicitly accounting for temporal visibility and adapting thresholds, the approach successfully resolves the blurriness and sparse reconstruction issues common in previous dynamic 3DGS implementations.