ToCo-Mesh

Preprint

ToCo-Mesh

Topology-Consistent Dynamic Mesh Reconstruction via Adaptive Tessellation and Surface-Aligned 2DGS

Adaptive where details emerge. Consistent through time.

Chuanjin Fan Wenjie Chang Aibing Li Bingzhou Wang Wenfei Yang Tianzhu Zhang

University of Science and Technology of China

Paper arXiv Watch video Code (Coming soon)
Overview of ToCo-Mesh, from topology-adaptive geometry optimization to surface-aligned 2D Gaussian splatting and high-fidelity reconstruction.
ToCo-Mesh reconstructs topology-consistent dynamic meshes from temporal images by adaptively refining a shared canonical template. Error-driven split and merge operations allocate mesh resolution to evolving details and propagate consistently across the sequence. Surface-aligned 2DGS are anchored to mesh faces, while their rendered normals guide geometry refinement for high-fidelity reconstruction.

Overview

Fine geometry without losing correspondence.

Per-frame extraction can recover local detail, but changes mesh connectivity over time. Fixed templates preserve correspondence, yet cannot allocate new resolution as details emerge. ToCo-Mesh refines a canonical template and propagates every topological operation to all timestamps, then uses surface-aligned 2DGS normals to improve the final geometry.

01

Consistent correspondence

Shared vertex identities and face connectivity across the sequence.

02

Adaptive tessellation

Split and merge operations place mesh resolution where details require it.

03

Surface-aligned 2DGS

Mesh-anchored Gaussians provide appearance and normal-guided refinement.

Interactive results

Three focused demonstrations will make topology consistency directly observable.

01

Correspondence

Track a Vertex Through Time

Temporal topology consistency is a core advantage of ToCo-Mesh: all frames share vertex identities and face connectivity as geometry deforms. Select a point to verify it directly—the highlighted vertex and one-ring neighborhood remain the same surface entities throughout all 200 frames.

Preparing viewer The mesh loads when this section enters view.
Select a vertex Click the mesh to reveal its identity through time.
Tracked vertex — No vertex selected
Vertex One-ring patch Trajectory

Drag to orbit Scroll to zoom

Frame 000 / 199
02

Shared UV editing

Paint Once, Move Everywhere

After reconstruction, we post-train a single shared UV texture on the topology-consistent mesh sequence produced by ToCo-Mesh. Paint on one pose below: the common surface parameterization carries the edit to the corresponding region in every frame, without retraining the reconstruction.

Paint onceEditable pose · Frame 003
Move everywhereAnimated pose · Frame 000
Live shared atlas
Draw on the left Duck The animated pose updates instantly.
Preparing shared UV viewer The Duck sequence loads when this section enters view.

Paint mode

Frame 000 / 199

One texture is shared by both poses.

03

Visual comparison

Reconstruction Visualization Results

Compare mesh geometry and Gaussian rendering under two fixed viewpoints per D-NeRF scene. Every active slider shows the same camera and temporal frame on both sides, making geometry and appearance directly comparable across methods.

Drag any active divider to compare the synchronized videos.

Baseline ToCo-Mesh
Fixed viewpoint

Switches mesh and rendering panels together.

Baseline 01 4D-GS
Baseline 02 SC-GS
Baseline 03 D-2DGS
Baseline 04 MaGS
Hellwarrior Mesh
Loading mesh comparison

The fixed-view video is ready for interactive playback.

Loading mesh comparison

The fixed-view video is ready for interactive playback.

Loading mesh comparison

The fixed-view video is ready for interactive playback.

Loading mesh comparison

The fixed-view video is ready for interactive playback.

Hellwarrior Render
Loading videos
Loading videos
Loading videos
Loading render comparison

The fixed-view video is ready for interactive playback.

Hook Mesh
Loading mesh comparison

The fixed-view video is ready for interactive playback.

Loading mesh comparison

The fixed-view video is ready for interactive playback.

Loading mesh comparison

The fixed-view video is ready for interactive playback.

Loading mesh comparison

The fixed-view video is ready for interactive playback.

Hook Render
Loading videos
Loading videos
Loading videos
Loading render comparison

The fixed-view video is ready for interactive playback.

Method

ToCo-Mesh separates discrete topology adaptation from continuous geometry and appearance refinement, forming a stable coarse-to-fine reconstruction pipeline.

Stage I

Topology-Adaptive Geometry Optimization

A canonical template is bound to temporal guide meshes. Error-driven split and merge operations refine this shared template and remain consistent across time.

Stage II

Texture Optimization and Mesh Refinement

Surface-aligned 2D Gaussians are anchored to mesh faces. Their rendered normals guide fine geometry updates while preserving high-quality appearance.

The two-stage ToCo-Mesh pipeline: topology-adaptive geometry optimization followed by texture optimization and normal-guided mesh refinement.
ToCo-Mesh pipeline. Select the figure to view it at full resolution.

Quantitative results

Geometry comparison on the DG-Mesh dataset

Per-scene Chamfer Distance (CD) and Earth Mover's Distance (EMD); lower is better.

Best Second Third
DG-Mesh dataset quantitative comparison
Type Method Corresp. Duck Horse Bird Beagle Torus2sphere Girlwalk
CD ↓EMD ↓ CD ↓EMD ↓ CD ↓EMD ↓ CD ↓EMD ↓ CD ↓EMD ↓ CD ↓EMD ↓
NeRF-Based D-NeRF × 0.9340.073 1.6850.280 1.5320.163 1.0010.149 1.7600.250 0.6010.190
TiNeuVox-B × 0.9690.059 1.9180.246 8.2640.215 0.8740.129 2.1150.203 0.5680.184
3DGS-Based 4D-GS × 1.1340.111 1.5000.272 2.3110.187 0.6440.106 2.1880.261 0.5960.315
Deformable-GS × 2.3660.115 1.5100.217 1.3580.141 1.1540.161 2.2100.248 1.1030.183
SC-GS × 1.3060.097 0.8970.177 0.8970.166 3.3590.147 2.0450.225 0.6230.203
Mesh-GS Based DG-Mesh × 0.7820.047 0.2970.164 0.5100.125 0.6230.114 1.5720.177 0.3980.151
D-2DGS × 1.1740.120 0.3910.177 0.3280.110 0.6160.110 1.5550.203 0.3160.133
MaGS ✓ 1.3500.053 1.9500.160 0.9920.086 1.4400.114 3.3170.123 1.2680.114
Ours ✓ 0.6890.050 0.3740.162 0.3870.084 0.5480.111 1.6040.121 0.3510.118

Color indicates rank within each scene and metric. Lower is better. Corresp. indicates whether a method provides explicit vertex-wise correspondence across frames.

Video

The full presentation covers the motivation, topology-adaptive optimization, surface-aligned 2DGS refinement, and dynamic reconstruction results.

Citation

Cite ToCo-Mesh.

Please use the following preprint entry for now. 

@misc{fan2026tocomesh,
  title  = {{ToCo-Mesh}: Topology-Consistent Dynamic Mesh Reconstruction via Adaptive Tessellation and {Surface-Aligned 2DGS}},
  author = {Fan, Chuanjin and Chang, Wenjie and Li, Aibing and Wang, Bingzhou and Yang, Wenfei and Zhang, Tianzhu},
  year   = {2026},
  eprint = {2609.29529},
  archivePrefix = {arXiv}
}