Research map / 3D vision / Depth and stereo
Practical Measurements of Translucent Materials with Inter-Pixel Translucency Prior
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 IPTNet, a multi-encoder CNN that recovers scattering parameters of homogeneous translucent materials from ordinary uncalibrated photos, guided by an inter-pixel translucency prior (translucent curves in RGB space).
Problem
Acquiring bulk scattering properties of translucent materials from ordinary photographs taken under unknown lighting and camera parameters, where differentiable rendering is unstable and dedicated hardware is costly.
Approach
Pixels at the same position across multiple differently lit images form curves through the origin in RGB space whose shape encodes translucency; this yields a pseudo-albedo map and a bias map fed to a multi-branch network with translucency-aware max-pooling feature fusion and a cascaded decoder that predicts Λ' before {σt', g}.
Key results
- On TrView, IPTNet reaches MSE 0.001 for Λ' and PSNR 38.68 on hexahedron versus 34.88 for ITN(M)-style baselines (ITN(S)).
- IPTNet predicts in about 30 ms, while the differentiable rendering optimization takes about 20 minutes to converge.
- Ablation: removing the inter-pixel prior (-pr) raises TrShape σt' MSE from 0.026 to 0.038 and lowers PSNR on the duck model from 35.96 to 33.92.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IPTNet instead of ITN(M) | Translucent material acquisition · TrView, hexahedron | PSNR: 35.94 -> 38.68 | Helped |
| Use IPTNet instead of ITN(M) | Translucent material acquisition · TrShape, hexahedron | PSNR: 34.58 -> 37.41 | Helped |
| Use IPTNet instead of ITN(M) | Translucent material acquisition · TrLight, hexahedron | PSNR: 29.62 -> 36.34 | Helped |
| Replace max pooling with mean pooling (ablation) | IPTNet · TrShape, hexahedron | PSNR: 37.41 -> 35.53 | Hurt |
| Replace max pooling with mean pooling (ablation) | IPTNet · TrLight, hexahedron | PSNR: 36.34 -> 34.63 | Hurt |
| Remove the pseudo-albedo and bias maps (ablation) | IPTNet · TrShape, hexahedron | PSNR: 37.41 -> 35.53 | Hurt |
| Remove the pseudo-albedo and bias maps (ablation) | IPTNet · TrLight, hexahedron | PSNR: 36.34 -> 31.74 | Hurt |
| Replace the cascaded decoder with a single decoder (ablation) | IPTNet · TrShape, hexahedron | PSNR: 37.41 -> 36.35 | Hurt |
| Replace the cascaded decoder with a single decoder (ablation) | IPTNet · TrLight, hexahedron | PSNR: 36.34 -> 32.92 | Hurt |
Surprising findings
- Differentiable rendering (Mitsuba 3 Dr) remains hard to converge even when lighting and geometry are provided, sometimes converging to incorrect results.
Limitations
- Pseudo-albedo/bias regression requires at least three input images.
- Inter-pixel prior fails for near-transparent objects that mix color with the background.
- Over-saturated pixels from specular materials can cause failure.
Experimental setup
Datasets: synthetic translucent renderings (Mitsuba) TrView TrLight TrShape
Benchmarks: TrView TrShape TrLight
Metrics: MSE over normalized parameters PSNR SSIM
Baselines: ITN(S) ITN(M) ITN(-Dr) Mitsuba 3 differentiable rendering (Dr) multi-view stereo (MVS) reconstruction
Builds on: Inverse Transport Network (ITN) Mitsuba 3 Henyey-Greenstein phase function principle of similarity
Closest papers by meaning
- Material Palette: Extraction of Materials from a Single Image (CVPR 2024 · similarity 0.86)
- Alchemist: Parametric Control of Material Properties with Diffusion Models (CVPR 2024 · similarity 0.86)
- Neural LightRig: Unlocking Accurate Object Normal and Material Estimation with Multi-Light Diffusion (CVPR 2025 · similarity 0.86)
- Eclipse: Disambiguating Illumination and Materials using Unintended Shadows (CVPR 2024 · similarity 0.85)
- OMG: Opacity Matters in Material Modeling with Gaussian Splatting (ICLR 2025 · similarity 0.85)
- Intrinsic Image Diffusion for Indoor Single-view Material Estimation (CVPR 2024 · similarity 0.85)
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.