Research map / 3D vision / Depth and stereo
Spin-UP: Spin Light for Natural Light Uncalibrated Photometric Stereo
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 Spin-UP, an unsupervised natural-light uncalibrated photometric stereo method that captures images on a rotating platform, initializes environment light from occluding boundaries, and optimizes neural inverse rendering with interval sampling and shrinking range computing.
Problem
Natural light uncalibrated photometric stereo (NaUPS): recovering surface normals, environment light, and reflectance from images under unknown natural lighting, which is ill-posed and highly ambiguous.
Approach
A spin-light capture setup (object rotating on a platform under static environment light) reduces light unknowns to a single uniform environment light plus per-image rotation angles; a boundary-pixel-based light initialization with diffuse/chromatic filters provides a prior, and training uses interval sampling and shrinking range computing for efficiency.
Key results
- On the synthetic dataset, Spin-UP achieves average normal MAE of 7.85 across groups (best in Table 4), versus 11.46 for S23 with point+environment light setting comparisons.
- Spin-UP reaches average MAE 8.72 on shape group vs 12.39 for S23 and 16.63 for S22 in Table 2.
- Interval sampling cuts training time from about 60 to 25 minutes per object and GPU memory from about 25 GB to 5 GB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Spin-UP instead of S23 | NaUPS normal estimation · Shape group (synthetic) | MAE: 12.39 -> 8.72 | Helped |
| use Spin-UP instead of S23 | NaUPS normal estimation · Light group (synthetic) | MAE: 10.84 -> 4.54 | Helped |
| use Spin-UP instead of S23 | NaUPS normal estimation · Reflectance group (synthetic) | MAE: 13.03 -> 9.08 | Helped |
| use Spin-UP instead of S23 | NaUPS normal estimation · Spatially varying material group (synthetic) | MAE: 13.53 -> 9.15 | Helped |
| add boundary-based light initialization instead of random initialization (ablation) | Spin-UP · Synthetic dataset | PU-PSNR: 16.86 -> 22.06 | Helped |
| add boundary-based light initialization instead of Fibonacci lattice initialization (ablation) | Spin-UP · Synthetic dataset | PU-SSIM: 0.52 -> 0.60 | Helped |
| add interval sampling (ablation) | Spin-UP · Spatially varying material group (synthetic) | MAE: 10.12 -> 9.15 | Helped |
| add shrinking range computing (ablation) | Spin-UP · Shape group (synthetic) | MAE: 10.10 -> 8.72 | Helped |
| add chromatic filter Fc (ablation) | Spin-UP · Synthetic dataset (average across groups) | MAE: 8.17 -> 7.85 | Helped |
| add low-pass spherical harmonics diffuse filter Fd SH (ablation) | Spin-UP · Spatially varying material group (synthetic) | MAE: 15.73 -> 9.15 | Helped |
| add threshold diffuse filter Fd TH (ablation) | Spin-UP · Spatially varying material group (synthetic) | MAE: 12.49 -> 9.15 | Helped |
Surprising findings
- The estimated environment light is noticeably influenced by object material, producing artifacts on objects with nearly uniform base color or spatially varying material.
Limitations
- Assumes infinitely far light sources, omitting spatially varying lighting.
- Base color of materials biases the estimated environment light; isotropic reflectance only, ignoring inter-reflections and anisotropy.
- Shadows are not computed iteratively, causing artifacts on complex shapes.
Experimental setup
Datasets: DiLiGenT-MV shapes Blender-rendered synthetic set with HDR environment maps self-captured real-world spin light dataset
Benchmarks: Spin-UP synthetic dataset (shape, light, reflectance, spatially varying material groups) Spin-UP real-world captured dataset
Metrics: mean angle error (MAE) PU-PSNR PU-SSIM
Baselines: HY19 S22 S23
Builds on: NeRF-style neural inverse rendering Disney BRDF model three-stage training schema (Ju et al.) spherical Gaussian light model Fibonacci lattice initialization
Closest papers by meaning
- A Benchmark Dataset and Evaluation for Non-Lambertian and Uncalibrated Photometric Stereo (CVPR 2016 · similarity 0.86)
- DANI-Net: Uncalibrated Photometric Stereo by Differentiable Shadow Handling, Anisotropic Reflectance Modeling, and Neural Inverse Rendering (CVPR 2023 · similarity 0.86)
- Uncalibrated Photometric Stereo Under Natural Illumination (CVPR 2018 · similarity 0.86)
- Sparse Views Near Light: A Practical Paradigm for Uncalibrated Point-light Photometric Stereo (CVPR 2024 · similarity 0.85)
- RNb-NeuS: Reflectance and Normal-based Multi-View 3D Reconstruction (CVPR 2024 · similarity 0.84)
- MetaGS: A Meta-Learned Gaussian-Phong Model for Out-of-Distribution 3D Scene Relighting (NeurIPS 2025 · similarity 0.83)
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.