Research map / 3D vision / Depth and stereo
What is Learned in Deep Uncalibrated Photometric Stereo?
Guanying Chen, Michael Waechter, Boxin Shi, Kwan-Yee K. Wong, Yasuyuki Matsushita
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LCNet with GCNet (guided by estimated normals and shading) | lighting estimation · DiLiGenT | light direction MAE (deg): 4.92 -> 3.32 | Improved |
| replace LCNet with GCNet | lighting estimation · Light Stage Data Gallery | light direction MAE (deg): 13.61 -> 9.20 | Improved |
| add end-to-end fine-tuning to L-Net1 + N-Net + L-Net2 (ablation) | GCNet · Synthetic Sphere | light direction error (deg): 2.52 -> 2.21 | Improved |
| add estimated normals as input to the second lighting stage (ablation) | GCNet · Synthetic Armadillo | light direction error (deg): 5.25 -> 3.92 | Improved |
| add N-Net and L-Net2 stages after L-Net1 (ablation) | GCNet · Synthetic Dragon | light direction error (deg): 5.80 -> 4.20 | Improved |
| use GCNet lightings instead of LCNet with IS18 | normal estimation · DiLiGenT | normal MAE (deg): 16.5 -> 8.6 | Improved |
| use GCNet lightings instead of LCNet with PS-FCN | normal estimation · DiLiGenT | normal MAE (deg): 9.7 -> 8.7 | Improved |
| use GCNet lightings instead of LCNet with ST14 | normal estimation · DiLiGenT | normal MAE (deg): 12.1 -> 10.8 | Improved |
| train LCNet on shading input instead of regular images (ablation) | LCNet · Synthetic Sphere | light direction error (deg): 3.03 -> 2.29 | Improved |
| train LCNet on attached shadows instead of regular images (ablation) | LCNet · Synthetic Dragon | light direction error (deg): 6.30 -> 9.78 | Hurt |
| train LCNet on combined shadows, specular and shading input (ablation) | LCNet · Synthetic Armadillo | light direction error (deg): 6.37 -> 2.12 | Improved |
| use GCNet under upward-biased lighting | GCNet vs LCNet + PS-FCN · Armadillo upward-biased | light direction error (deg): 7.10 -> 5.96 | Improved |
Closest papers by meaning
- Universal Photometric Stereo Network Using Global Lighting Contexts (CVPR 2022 · similarity 0.92)
- Self-Calibrating Photometric Stereo by Neural Inverse Rendering (ECCV 2022 · similarity 0.91)
- Deep Photometric Stereo for Non-Lambertian Surfaces (TPAMI 2020 · similarity 0.91)
- Deep Learning Methods for Calibrated Photometric Stereo and Beyond (TPAMI 2024 · similarity 0.91)
- Revisiting One-Stage Deep Uncalibrated Photometric Stereo via Fourier Embedding (TPAMI 2025 · similarity 0.91)
- Self-Calibrating Deep Photometric Stereo Networks (CVPR 2019 · similarity 0.91)
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.
