Research map / 3D vision / Face and shape
From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $\alpha$-NeuS
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.
Proves NeuS's density mapping is unbiased from fully transparent to fully opaque surfaces and extracts transparent and opaque surfaces jointly from the learned distance field using a DCUDF-based absolute-field projection.
Problem
Reconstructing both transparent (thin) and opaque objects simultaneously from multi-view images, which NeRF-family surface methods fail to do because they assume opaque surfaces and standard iso-surface extraction misses transparent surfaces.
Approach
Theoretically shows that NeuS's SDF-to-density mapping is unbiased across opacities: transparent surfaces correspond to non-negative local minima of the learned distance field and opaque ones to the zero iso-surface. It then extracts both by taking the absolute distance field and projecting a marching-cubes envelope onto its local minima with DCUDF's two-stage optimization.
Key results
- On the synthetic benchmark, the method reaches mean CD 5.67e-3 versus 23.41e-3 for NeuS at iso=0 and 7.71e-3 for NeuS at iso=0.005.
- The extracted surfaces achieve 100% completeness on synthetic scenes, while zero iso-surface extraction leaves roughly 20% to 40% holes.
- Qualitative results show transparent and opaque surfaces reconstructed on real-world captured scenes where NeuS with zero iso-surface leaves artifacts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use α-NeuS surface extraction instead of NeuS extraction at iso=0.005 | NeuS · Synthetic dataset mean | CD (×10^-3): 7.71 -> 5.67 | Helped |
| Use α-NeuS surface extraction instead of NeuS extraction at iso=0 | NeuS · Synthetic dataset mean | CD (×10^-3): 23.41 -> 5.67 | Helped |
Surprising findings
- Vanilla NeuS already learns transparent surfaces, but standard marching cubes at the zero iso-surface cannot extract them because their distance minima are positive.
Limitations
- Fails on scenes with heavy reflection or refraction, where NeuS distance fields become ambiguous.
- Targets thin transparent objects with little refraction; relies on polarizers to reduce reflection in real captures.
- Evaluation benchmark is small (5 synthetic and 5 real scenes).
Experimental setup
Datasets: synthetic Blender scenes (5) real-world captured scenes (5) DTU
Benchmarks: alpha-NeuS synthetic benchmark DTU
Metrics: Chamfer distance g2d d2g completeness percentage
Baselines: NeuS NeUDF alphaSurf
Builds on: NeuS DCUDF HF-NeuS NeUDF
Closest papers by meaning
- Opti-NeuS: Neural Reconstruction for Dual-Layered Transparent and Opaque Objects (CVPR 2026 · similarity 0.92)
- Differentiable Neural Surface Refinement for Modeling Transparent Objects (CVPR 2024 · similarity 0.90)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.87)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.87)
- Factored-NeuS: Reconstructing Surfaces, Illumination, and Materials of Possibly Glossy Objects (CVPR 2025 · similarity 0.87)
- Structure from Collision (CVPR 2025 · similarity 0.87)
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.