IMPLICIT 4D GAUSSIAN SPLATTING FOR FAST MOTION WITH LARGE INTER-FRAME DISPLACEMENTS

Existing 4D Gaussian Splatting (4DGS) methods struggle to reconstruct scenes with fast-moving objects and large inter-frame displacements. These approaches often fail because they either rely on canonical space deformations that omit fast-moving objects or use explicit 4D parameterizations that suffer from cross-frame interference, leading to blurred or vanished objects. This paper asks how to maintain high-fidelity Gaussian attributes for dynamic objects under these challenging motion conditions.

The authors introduce SPIN-4DGS (Spatiotemporal Position Implicit Network for 4DGS). Instead of explicitly optimizing Gaussian attributes (color, scale, rotation, opacity) for every frame—which is memory-intensive and prone to cross-frame interference—the authors use a two-stage process. First, they estimate and refine explicit spatiotemporal positions (x, y, z, t) for the Gaussians. Second, they use a lightweight feed-forward network to predict Gaussian attributes directly from these spatiotemporal coordinates. By using a 4D hash encoder and a multi-branch decoder, the model learns a shared implicit representation that ensures consistency across time while significantly reducing memory overhead.

SPIN-4DGS demonstrates superior performance on the CMU Panoptic Sports dataset, which contains rapid human motion and small, fast-moving objects. The method consistently outperforms strong baselines like D3DGS, achieving a notable +1.83 dB improvement in PSNR on the Basketball scene. Qualitative results show that while prior methods often blur or lose fast-moving objects, SPIN-4DGS produces sharp, stable, and accurate reconstructions. The authors also show that their framework can improve results even when reusing pre-trained Gaussian positions from other methods, suggesting that the implicit attribute prediction is the primary driver of the performance gain.