Research map / 3D vision / Face and shape
High-Fidelity Polarimetric Implicit 3D Reconstruction with View-Dependent Physical Representation
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 a polarimetric neural implicit surface reconstruction method that predicts incident Stokes vectors with MLPs, models specular/diffuse/transmission via Mueller matrices, and adds a view-dependent confidence map and AoLP fusion to reconstruct transparent and reflective objects.
Problem
Neural implicit 3D reconstruction assumes view-independent appearance, failing on transparent and highly reflective objects; prior polarimetric methods assume unpolarized incident light or only handle reflection.
Approach
Combines an SDF-based implicit surface with MLPs predicting polarized incident Stokes vectors and Fresnel/Mueller-matrix physical rendering, plus a view-dependent confidence map (from color variance across visible cameras) that down-weights view-dependent regions and a fused AoLP constraint weighted by ray-traced reflection percentage.
Key results
- On SMVP3D, average Chamfer Distance 1.43 vs. 1.52 for NeRSP (best baseline among those that run on all scenes).
- Average mean angular error on SMVP3D normals is 2.59 vs. 3.37 for NeRSP and 4.15 for PIR.
- Ablation on hedgehog normals: MAE drops from 1.68 (baseline) to 1.55 with view-dependent confidence and 1.42 with full model including polarization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed method instead of the strongest baseline | Ours vs. NeRSP · SMVP3D synthetic — Average | Chamfer Distance (CD): 1.52 -> 1.43 | Helped |
| Use the proposed method instead of the strongest baseline | Ours vs. NeRSP · SMVP3D synthetic — Dragon | Chamfer Distance (CD): 0.98 -> 0.94 | Helped |
| Use the proposed method instead of the strongest baseline | Ours vs. NeRSP · SMVP3D synthetic — Hedgehog | Chamfer Distance (CD): 1.38 -> 1.46 | Hurt |
| Use the proposed method instead of the strongest baseline | Ours vs. NeRSP · SMVP3D synthetic — David | Chamfer Distance (CD): 2.21 -> 1.88 | Helped |
| Use the proposed method instead of the strongest baseline | Ours vs. PIR · SMVP3D synthetic — Average | Mean angular error (MAE): 4.15 -> 2.59 | Helped |
| Add view-dependent confidence (VDC) (ablation) | Proposed method · SMVP3D synthetic — Hedgehog | Mean angular error (MAE): 1.68 -> 1.55 | Helped |
| Add the polarization module (ablation) | Proposed method · SMVP3D synthetic — Hedgehog | Mean angular error (MAE): 1.68 -> 1.48 | Helped |
| Add the polarization module to the VDC variant (ablation) | Proposed method · SMVP3D synthetic — Hedgehog | Mean angular error (MAE): 1.55 -> 1.42 | Helped |
| Add VDC to the polarization-only variant (ablation) | Proposed method · SMVP3D synthetic — Hedgehog | Mean angular error (MAE): 1.48 -> 1.42 | Helped |
Surprising findings
- NeISF failed to produce reasonable geometry on the dragon scene and had a very high average CD (11.44), indicating instability on some objects.
Limitations
- Performance degrades when the target object is fully transparent, since the view-dependent detection becomes less effective.
- Relies on captured polarized images with calibrated poses (Colmap) and masks from EiSeg labeling.
- Evaluated on only one synthetic dataset and a self-captured real set with limited object count.
Experimental setup
Datasets: SMVP3D self-captured real polarized object dataset
Benchmarks: SMVP3D
Metrics: Chamfer Distance mean angular error visual comparison
Baselines: PANDORA PIR NeRSP NeISF Ours-no-pol
Builds on: NeuS IDR NeISF PIR PANDORA
Closest papers by meaning
- NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images (CVPR 2024 · similarity 0.88)
- Glossy Object Reconstruction with Cost-effective Polarized Acquisition (CVPR 2025 · similarity 0.88)
- GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors (ICLR 2024 · similarity 0.87)
- NeISF: Neural Incident Stokes Field for Geometry and Material Estimation (CVPR 2024 · similarity 0.85)
- Differentiable Neural Surface Refinement for Modeling Transparent Objects (CVPR 2024 · similarity 0.85)
- PhyRecon: Physically Plausible Neural Scene Reconstruction (NeurIPS 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.