Integrating Multi-view Multi-light Surface Reconstruction into Cultural Heritage Workflows
Baptiste Brument, Robin Bruneau, Benjamin Coupry, Vincent Demoulin, Jean Mélou, Antoine Laurent, Fabien Castan, Jean-Denis Durou, et al.
cs.CV
Quick verdict
Systems paper wiring MVPS methods into Meshroom; four qualitative heritage case studies show fine-relief recovery, but evaluation is visual only, no metrics.
Summary
This systems paper integrates state-of-the-art multi-view photometric stereo (MVPS) into Meshroom, an open-source photogrammetry framework widely used in cultural heritage. Rather than proposing a new algorithm, the authors assemble a photometric-stereo ecosystem of interchangeable nodes (calibrated PS with automatic sphere detection, self-calibrated near-light PS, and the universal neural methods UniMS-PS, SDM-UniPS, LINO-UniPS), SAM 3-based automatic object masking, and a multi-view normal-and-reflectance integration node (RNb-NeuS2, plus their MPL-2.0 reimplementation OpenRNb for commercial use). The pipeline turns illumination-varying image sets — already captured for RTI with domes or hand-held flashes — into a single mesh textured with shading-free reflectance. Validation is qualitative on four real archaeological objects (terracotta head, engraved bone, flint biface, ceramic head), each compared against up to three photogrammetry packages and, for two objects, structured-light scans. The claimed outcome is markedly finer relief (beard strands, shallow engravings, flake scars, ceramic micro-granularity) than photogrammetry at a comparable image budget, with grazing-light relighting used to argue the relief is genuine. The only quantitative result is on DiLiGenT-MV: OpenRNb's mean Chamfer distance (0.24/0.30 mm) trails RNb-NeuS2 (0.18/0.22 mm). Main limitations: no quantitative evaluation on the heritage objects, no ablations over view or light counts, only two of the five photometric-stereo nodes exercised in the case studies, and NVIDIA-GPU dependency. Code and pipelines are released open-source.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 1
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A complete photometric-stereo ecosystem (calibrated, self-calibrated near-light, universal) plus SAM 3 masking and normal-and-reflectance integration can be assembled as Meshroom nodes into an operational end-to-end MVPS pipeline for heritage practitioners.
- On four real archaeological objects, the pipeline recovers fine relief (beard strands, shallow engravings, flake scars, ceramic micro-granularity) that photogrammetry at a comparable image budget does not resolve.
- The resulting reflectance texture is illumination-neutral (free of baked-in shading), unlike photogrammetric texturing (Fig. 2).
- OpenRNb, the MPL-2.0 reimplementation, approximates RNb-NeuS2 quality on DiLiGenT-MV (mean Chamfer 0.24 vs 0.18 mm on UniMS-PS normals; 0.30 vs 0.22 mm on SDM-UniPS), enabling commercial use.
- The full chain from raw multi-light images to a textured mesh runs in a few hours per object on a single GPU, comparable to conventional photogrammetry.
Strengths
Real integration gap addressed with the right vehicle
The paper identifies that MVPS research code exists but is 'unpackaged' while heritage practitioners already capture multi-light data for RTI (Sec. 1); grafting the chain onto Meshroom/AliceVision — with RAW/HDR ingestion, render-farm dispatch (Sec. 3.2) — is the adoption path the community actually uses, and all nodes plus ready-made pipelines are open-sourced (C3).
Genuinely modular PS ecosystem spanning lighting regimes
Five interchangeable nodes behind one interface: calibrated PS with automatic SphereDetection + LightingCalibration (Fig. 5), self-calibrated near-light PS fed by the MVS proxy [13], and universal methods UniMS-PS/SDM-UniPS/LINO-UniPS [22,23,11], covering calibrated to fully uncalibrated setups (Sec. 3.2).
Thoughtful anti-hallucination checks in the case studies
License-aware engineering with a commercial path
Practical performance story
Weaknesses
Entirely qualitative evaluation of the showcase results
The node ecosystem is under-exercised and the framing sentence is inconsistent
No ablations on the variables practitioners control
Sparse quantitative support overall
Baseline comparison design is loose
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is explicitly a **systems-integration paper**, not an algorithmic contribution — the Abstract says so directly ("Rather than proposing a new reconstruction algorithm"). The deliverables are (C1) a photometric-stereo node ecosystem in Meshroom spanning calibrated (SphereDetection + LightingCalibration, Fig. 5), self-calibrated near-light [13], and universal neural regimes (UniMS-PS [22], SDM-UniPS [23], LINO-UniPS [11]); (C2) SAM 3-based automatic masking plus a normal-and-reflectance integration node (RNb-NeuS2 [6], with the authors' MPL-2.0 reimplementation OpenRNb [41]); (C3) an open-source release of nodes and pipelines; and (C4) a qualitative validation on four archaeological objects. Most claims are appropriately scoped: the Introduction frames the gap accurately (MVPS lives in "unpackaged" research code while practitioners already capture multi-light data for RTI), and Sec. 5 candidly concedes the evaluation is qualitative. One claim does not survive reading: Sec. 4 states each object "exercises a different instantiation of the pipeline, namely a different photometric stereo method," yet Secs. 4.1, 4.3 and 4.4 all use the same self-calibrated near-light PS [13]; only 4.2 (engraved bone) uses UniMS-PS.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is multi-view photometric stereo (MVPS)?
MVPS photographs an object from several viewpoints, each under many different light directions. Photometric stereo converts each viewpoint's image stack into per-pixel surface normals (and often reflectance), and a fusion step integrates these normal maps across viewpoints into one 3D mesh. It recovers finer relief than standard photogrammetry on weakly textured objects such as stone, bone or ceramic.
Is the software from this paper available, and can it be used commercially?
Yes. The photometric-stereo nodes, masking and ready-made pipelines are released as open-source Meshroom plugins (the meshroomHub GitHub repositories). The reference integration backend RNb-NeuS2 carries a non-commercial NVIDIA research license, but the authors also provide OpenRNb, an MPL-2.0 reimplementation that permits commercial use, at some accuracy cost on the DiLiGenT-MV benchmark (mean Chamfer 0.24 vs 0.18 mm in one tested configuration).
Does this pipeline replace photogrammetry for digitising artefacts?
Not exactly. The paper positions it as a complement: photogrammetry remains strong for global shape, while the MVPS pipeline better recovers fine surface relief and produces shading-free textures. In the four reported case studies it resolved engravings and micro-relief that photogrammetry missed, but the comparison is qualitative and requires multi-light capture (RTI dome, turntable, or hand-held flash) plus an NVIDIA GPU.
What equipment do I need to use this pipeline?
You need multi-light image sets: an RTI/LED dome, a camera-plus-turntable rig, or a hand-held flash. The reported experiments used a 45.7 MP full-frame camera with 8-23 viewpoints under 12-35 illuminations each. Some stages (universal photometric-stereo networks, parts of integration) require an NVIDIA/CUDA GPU; the full chain runs a few hours per object.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are not analyzable from the extracted text; all visual-quality claims (Figs. 1, 6-9) are judged from captions and prose only, which is where most of the paper's evidence lives.
- Per-stage runtimes, hardware details, and any appendix content may exist in figure panels not captured by the text extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.