Research map / 3D vision / NeRF and radiance fields
VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources
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 VMINer, a multi-view inverse rendering framework that jointly uses far-field (spherical Gaussian) and near-field (point light) lighting under varied capture conditions, combining NeRF-style radiance rendering with physically-based surface rendering on SDF and material fields.
Problem
Inverse rendering from multi-view images is severely ambiguous between lighting and material; prior methods either assume one fixed lighting or model only far-field lights, and flashlight-based methods discard ambient observations.
Approach
A lightweight versatile lighting model that represents far-field lights as 128-lobe spherical Gaussians and near-field lights as point lights with adjustable position and SH anisotropy, combined with per-lighting neural radiance and physically-based rendering plus a self-consistency loss between them.
Key results
- With single far-field light plus flashlight, VMINer reaches 10.89 MAngE normal error and 30.75 PSNR FV relit, versus WildLight's 11.49 and 27.71, while training in 60 vs 1440 minutes.
- With single far-field light only, VMINer achieves 12.39 normal MAngE, comparable to TensoIR (17.66) and NVDiffRecMC (16.24) on albedo and relit metrics.
- Adding a flashlight to single far-field input greatly improves albedo PSNR from 28.20 to 32.09 (own method).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a flashlight to the single far-field-light capture | VMINer · 6 synthetic scenes | diffuse albedo PSNR: 24.50 -> 31.62 | Helped |
| add a flashlight to the single far-field-light capture | VMINer · 6 synthetic scenes | FV relit PSNR: 27.46 -> 32.00 | Helped |
| add a second far-field lighting condition | VMINer · 6 synthetic scenes | FV relit PSNR: 27.46 -> 29.27 | Helped |
| add a flashlight to the two-far-field-light capture | VMINer · 6 synthetic scenes | FV relit PSNR: 29.27 -> 32.38 | Helped |
| remove the self-consistency loss Lcon (ablation) | VMINer · 6 synthetic scenes, far-field + flashlight | albedo PSNR: 27.70 -> 28.05 | Hurt |
| omit explicit cosine weakening and incident-radiance multiplication from cnear (ablation) | VMINer · 6 synthetic scenes, far-field + flashlight | specular relit PSNR: 28.87 -> 29.36 | Hurt |
| replace encmat(x) with encgeo(x) (ablation) | VMINer · 6 synthetic scenes, far-field + flashlight | specular relit PSNR: 28.92 -> 29.36 | Hurt |
Surprising findings
- Adding a second far-field lighting gives smaller gains than adding a flashlight, and a further flashlight on top of two far-field lights yields only marginal improvement.
Limitations
- Does not model background, requiring foreground masks for each image.
- Assumes fixed gamma 2.2 tone mapping, ignoring unknown ISP curves.
- Relies on switching near-field lights on/off during capture to separate lighting contributions.
Experimental setup
Datasets: 6 synthetic scenes (rendered) real-world captures under ambient light and flashlight
Benchmarks: 6 synthetic scenes (authors' own)
Metrics: MAngE PSNR SSIM
Baselines: TensoIR NVDiffRecMC WildLight
Builds on: NeuS Instant-NSR WildLight TensoIR NeRFactor
Closest papers by meaning
- Neural Inverse Rendering from Propagating Light (CVPR 2025 · similarity 0.88)
- RobIR: Robust Inverse Rendering for High-Illumination Scenes (NeurIPS 2024 · similarity 0.88)
- IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images (CVPR 2025 · similarity 0.88)
- Inverse Rendering using Multi-Bounce Path Tracing and Reservoir Sampling (ICLR 2025 · similarity 0.87)
- ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images (CVPR 2024 · similarity 0.87)
- Inverse Rendering of Glossy Objects via the Neural Plenoptic Function and Radiance Fields (CVPR 2024 · similarity 0.87)
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.