Research map / 3D vision / Depth and stereo

Glossy Object Reconstruction with Cost-effective Polarized Acquisition

CVPR 2025 · 3D vision › Depth and stereo · approach: Non-line-of-sight imaging · 2 citations (OpenAlex)

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.

Reconstructs geometry and radiance of glossy objects from multi-view single-polarization-angle images captured with an RGB camera and a linear polarizer, using a polarimetric BRDF neural implicit model with the unknown polarizer angle optimized jointly.

Problem

3D reconstruction and diffuse/specular separation for glossy objects from RGB images is ill-posed; existing polarization methods need expensive polarization cameras or precise polarizer calibration.

Approach

Attach a linear polarizer to an off-the-shelf RGB camera, capture one polarized image per view with unknown polarizer angle, and render polarized images end-to-end from a VolSDF/Ref-NeRF-style neural implicit model using pBRDF Stokes-vector physics, optimizing the polarizer angle jointly with the network.

Key results

• On the RedOx object, the method reaches 30.88 PSNR / 0.9774 SSIM and chamfer distance 2.23e-4, outperforming NVDiffRec, InvRender, and NeRO on both rendering and geometry.
• On the synthetic Bust model, the method is comparable to PANDORA (mixed PSNR 26.53 vs 26.86; normal MAE 4.227 vs 4.096 degrees).
• The estimated polarizer angle has error below 5 degrees on the tested inputs.

Measured effects reported in the paper

Surprising findings

• NVDiffRec achieves high PSNR/SSIM despite erroneous geometry and poor diffuse/specular separation, suggesting rendering metrics alone do not reflect geometry quality.

Limitations

• Relies on color-bleeding-prone decomposition; acknowledged challenges such as color bleeding
• Geometry ground truth for real objects requires invasive painting and scanning, limiting evaluation scale
• Synthetic comparison with PANDORA is only comparable, not superior

Experimental setup

Datasets: PANDORA Owl and Gnome raw polarization data rendered Bust model self-captured real objects (RedOx, GreenOx, Cat, Horse, Lays)

Benchmarks: self-captured glossy object set (RedOx, GreenOx, Cat, Horse, Lays) rendered Bust model PANDORA Owl and Gnome

Metrics: PSNR SSIM chamfer distance mean angular error

Baselines: NeuralPIL PhySG NVDiffRec InvRender Ref-NeuS NeRO PANDORA

Builds on: PANDORA VolSDF Ref-NeRF polarimetric BRDF model COLMAP

Closest papers by meaning

• PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · similarity 0.88)
• NeISF: Neural Incident Stokes Field for Geometry and Material Estimation (CVPR 2024 · similarity 0.88)
• High-Fidelity Polarimetric Implicit 3D Reconstruction with View-Dependent Physical Representation (AAAI 2025 · similarity 0.88)
• RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.88)
• NeISF++: Neural Incident Stokes Field for Polarized Inverse Rendering of Conductors and Dielectrics (CVPR 2025 · similarity 0.86)
• PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation (NeurIPS 2026 · similarity 0.86)