Research map / 3D vision / Depth and stereo
DiLiGenRT: A Photometric Stereo Dataset with Quantified Roughness and Translucency
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.
Introduces DiLiGenRT, a real-world photometric stereo dataset of 54 hemispheres/spheres with quantified roughness (9 levels) and translucency (6 levels), and benchmarks classical and learning-based methods on it.
Problem
Existing real-world photometric stereo datasets control reflectance with qualitative semantic labels (e.g., specular, steel), so it is hard to measure how reliable methods are under non-Lambertian reflectance, especially translucency.
Approach
Manufacture real objects with a shared sphere geometry whose roughness is set by sandblasting molds with 9 grit sizes and whose translucency is set by 6 silicone-pigment concentrations, measure both properties with an interferometer (Sa) and a Beer-Lambert film setup (sigma_t), and report MAE as performance profile matrices over these quantified reflectance levels.
Key results
- SDM-UniPS achieves the smallest average MAE on DiLiGenRT (3.61 mean / 3.69 median), about threefold smaller than the second-best MS-PS (10.09/8.61).
- Highly translucent surfaces remain hard, with MAEs generally exceeding 20 degrees for most methods; translucency has a larger effect than roughness.
- Fine-tuning PS-FCN on the synthetic PS-SSS dataset reduces its mean MAE from 16.38 to 9.92 degrees.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune PS-FCN on synthetic translucent PS-SSS data | PS-FCN · DiLiGenRT | MAE (mean, degrees): 16.38 -> 9.92 | Helped |
| Fine-tune PS-FCN on synthetic translucent PS-SSS data | PS-FCN · DiLiGenRT, high translucency (σt ≤ 0.57) | MAE | Helped |
| Use a position thresholding strategy | LS on DiLiGenRT · DiLiGenRT | MAE: 22.42/18.96 -> 15.65/5.48 | Helped |
| Use DeconvPS to address subsurface scattering | LS on translucent surfaces · DiLiGenRT | MAE | Helped |
| Increase surface translucency | Photometric stereo methods · DiLiGenRT | MAE | Hurt |
| Use all-pixel-based methods rather than per-pixel-based methods with 10 lights | Photometric stereo on DiLiGenRT · DiLiGenRT, 10 lights | MAE: 49 of 54 materials | Helped |
| Reduce the number of lights from 100 to 10 | Photometric stereo on opaque and semi-translucent surfaces · DiLiGenRT | MAE | Mixed |
| Evaluate SDM-UniPS on distinct, non-sphere shapes | SDM-UniPS · DiLiGenRT-S | MAE: 18.6; 29.5; 23.2 | Hurt |
Surprising findings
- SDM-UniPS performs well on DiLiGenRT spheres but its MAE rises sharply on the diverse shapes of DiLiGenRT-S, possibly from overfitting to sphere-like shapes in its training data.
- Good performance on opaque specular highlights does not necessarily transfer to translucent surfaces.
Limitations
- Highly translucent surfaces remain challenging for most existing methods, with errors above 20 degrees.
- Sphere geometry in DiLiGenRT shares one shape, so shape-related errors are isolated only in the main set; DiLiGenRT-S has just three extra shapes.
- Translucency is controlled by only 6 levels and shares IOR across objects, limiting coverage.
Experimental setup
Datasets: DiLiGenRT PS-SSS PS-Wild PS-Mix DiLiGenT DiLiGenT102
Benchmarks: DiLiGenRT DiLiGenRT-S
Metrics: mean angular error (degrees) mean MAE median MAE
Baselines: LS TH28 TH46 DeconvPS ST14 PF14 CNN-PS PS-FCN Att-PSN MS-PS SDPS-Net SDM-UniPS
Builds on: DiLiGenT102 CyclesPS Disney Principled BSDF
Closest papers by meaning
- A Benchmark Dataset and Evaluation for Non-Lambertian and Uncalibrated Photometric Stereo (CVPR 2016 · similarity 0.91)
- DiLiGenT102: A Photometric Stereo Benchmark Dataset With Controlled Shape and Material Variation (CVPR 2022 · similarity 0.91)
- Universal Photometric Stereo Network Using Global Lighting Contexts (CVPR 2022 · similarity 0.87)
- DANI-Net: Uncalibrated Photometric Stereo by Differentiable Shadow Handling, Anisotropic Reflectance Modeling, and Neural Inverse Rendering (CVPR 2023 · similarity 0.86)
- Differentiable Display Photometric Stereo (CVPR 2024 · similarity 0.84)
- RNb-NeuS: Reflectance and Normal-based Multi-View 3D Reconstruction (CVPR 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.