Research map / 3D vision / Depth and stereo
Sparse Views Near Light: A Practical Paradigm for Uncalibrated Point-light 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 an end-to-end uncalibrated multi-view point-light photometric stereo framework built on VolSDF-style volume rendering, handling ambient light and near LED point lights from sparse viewpoints without training data.
Problem
Multi-view photometric stereo typically needs dark rooms, calibrated or distant directional lights, dense viewpoints, and multi-stage pipelines, which limits practical 3D reconstruction from sparse, wide-baseline views of textureless objects.
Approach
Combines VolSDF volume rendering with a physically based point-light plus ambient-light model, jointly optimizing SDF geometry, uncalibrated light positions and BRDF parameters end-to-end, with ambient subtraction, a robust L1 loss with brightest-view selection against cast shadows, and a closed-form IRLS solution for the diffuse albedo per pixel.
Key results
- On synthetic scans with 6 viewpoints (5 for dog), the method reports lower RMSE and MAE than [56], [10] directional/point-light and PS-NeRF [60]; e.g., dog1 RMSE 0.4 vs 2.1 for [56] and 2.1 for [60], MAE 4.6 vs 18.8 for [60].
- With only two wide-baseline viewpoints, the MLP-free albedo variant still outperforms PS-NeRF visually, avoiding the distortions seen in [60].
- Requires no pre-training or training data, and works with ambient light plus near LED lighting captured with a smartphone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use uncalibrated point-light photometric stereo instead of the compared reconstruction baselines | Ours · dog1 | RMSE×100: 2.1 -> 0.4 | Helped |
| use uncalibrated point-light photometric stereo instead of the compared reconstruction baselines | Ours · dog2 | RMSE×100: 2.1 -> 0.4 | Helped |
| use uncalibrated point-light photometric stereo instead of the compared reconstruction baselines | Ours · girl1 | RMSE×100: 0.9 -> 0.3 | Helped |
| use uncalibrated point-light photometric stereo instead of the compared reconstruction baselines | Ours · girl2 | RMSE×100: 1.5 -> 0.3 | Helped |
| use optimal diffuse albedo instead of an MLP diffuse-albedo network (ablation) | Ours · two viewpoints, wide baseline | 3D reconstruction quality | Helped |
| use optimal diffuse albedo instead of an MLP diffuse-albedo network (ablation) | Ours · sparse viewpoints | 3D reconstruction quality | Helped |
Surprising findings
- Calibrated PS baseline [10] with point lights performed worse than the directional version on several objects, e.g., girl1 RMSE 2.3 (point) vs 6.1 (directional).
Limitations
- Ablated two-view variant shows artifacts in some regions; wide-baseline two-view results remain imperfect.
- Does not explicitly model indirect lighting or cast shadows; relies on a truncation heuristic for shadows.
- Ambient subtraction doubles noise variance, yielding noisier inputs.
Experimental setup
Datasets: DiLiGenT-MV
Benchmarks: DiLiGenT-MV synthetic dog1/dog2/girl1/girl2 scans real-world bird/squirrel/hawk/rooster/flamingo/pumpkin scans
Metrics: RMSE (x100) vertex-to-mesh distance MAE angular normal error
Baselines: Wu et al. sparse-view neural reconstruction [56] [10] calibrated PS directional [10] calibrated PS point-light PS-NeRF [60] NeuS-style sparse view methods [33, 47, 65]
Builds on: VolSDF PS-NeRF Disney BRDF NeuS
Closest papers by meaning
- Fast Light-Weight Near-Field Photometric Stereo (CVPR 2022 · similarity 0.89)
- Universal Photometric Stereo Network Using Global Lighting Contexts (CVPR 2022 · similarity 0.88)
- TurboSL: Dense Accurate and Fast 3D by Neural Inverse Structured Light (CVPR 2024 · similarity 0.87)
- Generative Multiview Relighting for 3D Reconstruction under Extreme Illumination Variation (CVPR 2025 · similarity 0.86)
- RNb-NeuS: Reflectance and Normal-based Multi-View 3D Reconstruction (CVPR 2024 · similarity 0.86)
- MVCPS-NeuS: Multi-view Constrained Photometric Stereo for Neural Surface Reconstruction (CVPR 2024 · similarity 0.86)
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.