Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Dual Contouring
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Self-Supervised Dual Contouring (SDC), a neural meshing method trained with distance and normal-consistency losses on input SDFs instead of QEF-based supervision, and uses it as a regularizer for deep implicit networks and in end-to-end single-view reconstruction.
Problem
Learning-based isosurface extraction is typically trained supervised on axiomatically computed dual contouring meshes, inheriting their biases and artefacts (e.g., self-intersections from QEF vertex placement on noisy or sharp SDFs).
Approach
Extends Neural Dual Contouring with two self-supervised losses: a distance loss matching |SDF| at grid nodes to point-to-mesh distance, and a normal-consistency loss aligning SDF-gradient normals with quad face normals; the same mesh-based loss is also used as a regularizer for training deep implicit networks and jointly in single-view reconstruction.
Key results
- On ABC analytical SDFs, SDC reaches CD 3.3 and 9.7 self-intersecting triangles per mesh versus 3.6 and 43.1 for NDC.
- On ShapeNet DIN regularization, SDC improves 3D IoU to 86.0 versus 83.9 for CSDF without regularization and 84.8 for NDC regularization.
- In single-view reconstruction, the end-to-end SDC model achieves CD 9.1 and SSIM 88.5 versus 11.3 and 84.1 for MeshSDF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use self-supervised SDC instead of NDC for meshing | Neural Dual Contouring · ABC (analytical SDF) | CD: 3.6 -> 3.3 | Helped |
| use self-supervised SDC instead of NDC for meshing | Neural Dual Contouring · Thingi10K (analytical SDF) | CD: 4.0 -> 3.6 | Helped |
| use self-supervised SDC instead of NDC for meshing | Neural Dual Contouring · Thingi10K (NGLOD-predicted SDF) | CD: 4.7 -> 4.2 | Helped |
| add SDC mesh-based regularization | DeepSDF / CSDF · ShapeNet surface reconstruction | CD: 3.0 -> 2.6 | Helped |
| add Eikonal regularization | DeepSDF / CSDF · ShapeNet surface reconstruction | CD: 3.0 -> 3.2 | Hurt |
| use SDC rather than NDC for mesh-based regularization | DeepSDF / CSDF · ShapeNet surface reconstruction | CD: 2.9 -> 2.6 | Helped |
| jointly fine-tune the implicit model and mesher (ablation) | SDC single-view reconstruction · ShapeNet single-view reconstruction | CD: 9.9 -> 9.1 | Helped |
| add SDC mesh-based regularization | DeepSDF / CSDF · ShapeNet surface reconstruction | 3D IoU: 83.9 -> 86.0 | Helped |
| add SDC mesh-based regularization | DeepSDF / CSDF · ShapeNet surface reconstruction | NC: 78.1 -> 78.9 | Helped |
| add SDC mesh-based regularization | DeepSDF / CSDF · ShapeNet surface reconstruction | LSD-A: 3.1 -> 2.7 | Helped |
| jointly fine-tune the implicit model and mesher (ablation) | SDC single-view reconstruction · ShapeNet single-view reconstruction | NC: 72.0 -> 72.8 | Helped |
| jointly fine-tune the implicit model and mesher (ablation) | SDC single-view reconstruction · ShapeNet single-view reconstruction | SSIM: 86.4 -> 88.5 | Helped |
Surprising findings
- NDC-based regularization degrades DIN reconstruction relative to SDC, attributed to QEF being ill-defined for imperfect implicit functions.
- Marching Cubes and FlexiCubes produce zero self-intersections but have inferior geometric metrics.
Limitations
- Meshing and regularization operate on a regular grid, limiting mesh resolution.
- Normal consistency loss was unstable in the single-view setting with noisy evaluated SDFs and was dropped there.
- Strict manifoldness and self-intersection-free guarantees are not theoretically established.
Experimental setup
Datasets: ABC Thingi10K ShapeNet
Benchmarks: ABC Thingi10K ShapeNet
Metrics: Chamfer Distance Normal Consistency Self-Intersection Edge Chamfer Distance 3D-IoU LSD-P LSD-A SSIM
Baselines: Marching Cubes Dual Contouring Neural Marching Cubes Neural Dual Contouring FlexiCubes CSDF MeshSDF DISN
Builds on: Neural Dual Contouring Dual Contouring DeepSDF MeshSDF Curriculum DeepSDF
Closest papers by meaning
- Iso-Points: Optimizing Neural Implicit Surfaces With Hybrid Representations (CVPR 2021 · similarity 0.86)
- Analytic Marching: An Analytic Meshing Solution from Deep Implicit Surface Networks (ICML 2020 · similarity 0.86)
- Self-Supervised 3D Mesh Reconstruction From Single Images (CVPR 2021 · similarity 0.85)
- Faithful Contouring: Near-Lossless 3D Voxel Representation Free from Iso-surface (CVPR 2026 · similarity 0.85)
- Diffeomorphic Mesh Deformation via Efficient Optimal Transport for Cortical Surface Reconstruction (ICLR 2024 · similarity 0.84)
- Flatten Anything: Unsupervised Neural Surface Parameterization (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.