GauSTAR: Gaussian Surface Tracking and Reconstruction

TL;DR

GauSTAR tackles the challenge of dynamic 3D scene reconstruction with changing topology by binding Gaussians to mesh faces to produce Gaussian Surfaces that deliver photorealistic rendering and accurate geometry, while enabling robust tracking across topology changes. The method combines a scene-flow based initialization, fixed-topology Gaussian surface optimization, adaptive Gaussian unbinding to detach and spawn new surfaces, and re-meshing to update topology, all guided by RGB-D supervision and regularizers. A novel Gaussian unbinding mechanism and TSDF-based remeshing maintain temporal coherence and enable the reconstruction of newly visible surfaces, achieving state-of-the-art appearance and tracking metrics on multi-view data. GauSTAR’s integration of explicit geometry with Gaussian-based appearance supports versatile applications in VR/XR, telepresence, and beyond, while highlighting practical considerations and limitations for complex topology changes and non-LR-visible materials.

Abstract

3D Gaussian Splatting techniques have enabled efficient photo-realistic rendering of static scenes. Recent works have extended these approaches to support surface reconstruction and tracking. However, tracking dynamic surfaces with 3D Gaussians remains challenging due to complex topology changes, such as surfaces appearing, disappearing, or splitting. To address these challenges, we propose GauSTAR, a novel method that achieves photo-realistic rendering, accurate surface reconstruction, and reliable 3D tracking for general dynamic scenes with changing topology. Given multi-view captures as input, GauSTAR binds Gaussians to mesh faces to represent dynamic objects. For surfaces with consistent topology, GauSTAR maintains the mesh topology and tracks the meshes using Gaussians. For regions where topology changes, GauSTAR adaptively unbinds Gaussians from the mesh, enabling accurate registration and generation of new surfaces based on these optimized Gaussians. Additionally, we introduce a surface-based scene flow method that provides robust initialization for tracking between frames. Experiments demonstrate that our method effectively tracks and reconstructs dynamic surfaces, enabling a range of applications. Our project page with the code release is available at https://eth-ait.github.io/GauSTAR/.

Figures (6)

• Figure 1: We propose GauSTAR, a novel method that (a) enables photo-realistic rendering, surface reconstruction, and 3D tracking for dynamic scenes while handling topology changes. (b) GauSTAR adapts to topology changes through two mechanisms: consistent tracking for stable surfaces (red circles) and dynamic surface generation for newly appearing geometry (orange circles).
• Figure 2: Taking multi-view captures as input, GauSTAR tracks and reconstructs dynamic objects frame by frame. For each frame, GauSTAR first warps the previous frame's result using scene flow (\ref{['sec:scene_flow_warping']}). It then reconstructs Gaussian Surfaces (\ref{['sec:gaussian_surface']}) by fixed-topology reconstruction (\ref{['sec:basic_tracking']}). To handle topology-changing surfaces, GauSTAR detects topology changes, unbinds Gaussians on these surfaces, and adds new Gaussians as needed (\ref{['sec:Gaussian_unbinding']}). Finally, the Gaussian Surfaces are updated through re-meshing (\ref{['sec:surface_remeshing']}).
• Figure 3: Details of the mesh update process. (a) Visualization of unbinding weights defined in \ref{['eq:unbind_weight']}, where red indicates high weights in topology-changing regions. (b) Mesh connection process between original and new surfaces, with blue dotted lines showing vertex correspondences.
• Figure 4: Comparisons of appearance and geometry reconstruction. Dynamic 3D Gaussians luiten2024dynamic and PhysAvatar zheng2024physavatar yield suboptimal reconstruction results. HumanRF icsik2023humanrf and 2DGS huang20242d, lacking tracking capabilities, struggle under heavy occlusion. In contrast, GauSTAR provides high-quality reconstruction while supporting tracking. Additional comparisons are provided in our supplementary materials.
• Figure 5: Tracking comparisons using AprilTags. GauSTAR achieves more accurate tracking results, with predicted (red) and ground truth (blue) trajectories of tag centers shown.