TRiGS: Temporal Rigid-body Motion for Scalable 4D Gaussian Splatting

Recent advancements in 4D Gaussian Splatting (4DGS) have demonstrated impressive capabilities in real-time dynamic scene reconstruction. However, state-of-the-art methods typically rely on piecewise linear approximations or temporally constrained primitives, allowing Gaussians to simply appear and disappear at arbitrary times. This highly flexible but unconstrained approach leads to severe temporal fragmentation, where primitives are constantly destroyed and recreated. This rapid proliferation of Gaussians not only destroys the temporal identity of objects but also incurs prohibitive memory costs, making it nearly impossible to scale to extended video sequences.To address this fundamental limitation, we propose TRiGS, a novel 4D representation that integrates classical SE(3) kinematics to model temporally robust rigid-body motion. Instead of fragmenting motion into discontinuous linear segments, TRiGS optimizes continuous and physically grounded rigid-body transformations for Gaussian primitives. This enables individual primitives to maintain their long-term temporal identity. By preventing the motion approximation errors that cause unnecessary multiplication of Gaussians, our approach effectively mitigates unbounded memory growth. Extensive experiments demonstrate that TRiGS achieves high-fidelity rendering on standard benchmarks while uniquely scaling to extended video sequences (e.g., 600 to 1200 frames) without severe memory bottlenecks. By enforcing long-term kinematic consistency, our method significantly outperforms existing approaches in temporal stability for prolonged dynamic scenes.