Title: WebSpline: Leveraging Structure-Informed Splines for High-Speed 3D Gaussian Reconstruction from Single-View Video

Abstract:

Reconstructing dynamic scenes from monocular footage is a persistent hurdle in computer vision, primarily because current techniques often fail to maintain a balance between global structural integrity and localized detail when multi-view information is scarce. To overcome this limitation, we introduce WebSpline, a new dynamic 3D Gaussian framework designed to deliver high-fidelity, structurally coherent reconstructions from single-view videos while ensuring rapid rendering speeds.

Central to our approach is the Structure-Informed Spline (SIS) representation. This mechanism models the trajectory of each dynamic Gaussian using a learnable cubic Hermite spline, with motion patterns structurally organized through an auxiliary Structural Proxy Graph (SPG). Our optimization process unfolds in two distinct phases: First, the SPG is initialized using 2D point tracks and subsequently refined via temporal rigidity regularization, thereby establishing structural consistency for moving entities throughout the video sequence. Second, the SIS representation is derived from this refined SPG and further optimized by enforcing both spatial and structural neighborhood constraints.

During inference, Gaussian motion is determined exclusively by evaluating the learned SIS, which facilitates immediate rendering. We conducted extensive evaluations on rigorous monocular dynamic scene benchmarks, specifically the iPhone and NVIDIA datasets. The results indicate that WebSpline achieves state-of-the-art rendering quality, surpassing WorldTree—the second-best performer on the iPhone dataset—by delivering rendering speeds more than ten times faster.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC