WebSpline icon WebSpline

Structure-Informed Splines for Real-Time
3D Gaussians from Monocular Videos

Jongmin Park*      Jeonghwan Yun*      Minh-Quan Viet Bui      Munchurl Kim
*Co-first authors (equal contribution)
KAIST, South Korea
VICLab Logo Video and Image Computing Lab (VICLab)
arXiv Code

Visual Comparisons

We evaluate WebSpline on two widely used monocular video benchmarks: the iPhone and NVIDIA datasets. These visual comparisons demonstrate that our method produces novel view renderings with clearer boundaries and sharper details of dynamic objects compared to state-of-the-art approaches.

Gaussian Trajectory Visualization

To further evaluate our motion modeling, we visualize the learned Gaussian trajectories on the iPhone dataset. These comparisons illustrate how WebSpline smoothly and accurately captures the complex dynamics of moving objects over time compared to baseline methods.

Abstract

TL;DR: We propose WebSpline, a novel dynamic 3D Gaussian framework that couples a Structure-Informed Spline (SIS) representation with a Structural Proxy Graph (SPG), achieving state-of-the-art reconstruction quality from monocular videos while rendering over 10× faster than the second-best method on the iPhone dataset.

teaser

Framework Architecture

architecture

Overview of WebSpline. WebSpline models each dynamic Gaussian trajectory using the Structure-Informed Spline (SIS) representation, initialized from the Structural Proxy Graph (SPG). For SIS optimization, we define two types of neighborhoods for each dynamic Gaussian, spatial and structural, to enforce coherent spline motion while capturing fine-grained dynamics.