Research map / 3D vision / Depth and stereo
Polarization Wavefront Lidar: Learning Large Scene Reconstruction from Polarized Wavefronts
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.
Introduces a polarization wavefront lidar sensor (PolLidar) and a neural network that reconstructs distance, surface normals, and material properties from raw polarized time-resolved wavefronts for long-range automotive scenes.
Problem
Conventional lidar uses only time-of-flight intensity and ignores polarization, and prior polarization methods are limited to indoor/object-scale scenes or passive cameras that fail at night; accurate normals and distance for large outdoor scenes up to 100 m are needed.
Approach
A rotating-ellipsometry polarimetric lidar that records raw time-resolved wavefronts, with ellipsometric Mueller-matrix recovery and peak-based slicing feeding a TransUNet-style network that predicts normals and distance directly from the wavefronts.
Key results
- On synthetic data, normal mean angular error 8.71° vs 18.64° for PCA (53% reduction).
- Distance mean absolute error 19 cm vs 32 cm for argmax peak-finding on synthetic data (41% reduction).
- On real data, normal mean angular error 15.76° vs 18.75° for PCA; distance MAE 20 cm vs 24 cm (17% reduction).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed reconstruction method instead of PCA | Normal reconstruction · Synthetic data | mean angular error: 18.64 -> 8.71 (53%) | Helped |
| Use the proposed reconstruction method instead of PCA | Normal reconstruction · Experimental data | mean angular error: 18.75 -> 15.76 (16%) | Helped |
| Use raw wavefront reconstruction instead of conventional argmax peak-finding | Distance estimation · Synthetic data | mean absolute distance error: 32cm -> 19cm (41%) | Helped |
| Use raw wavefront reconstruction instead of conventional argmax peak-finding | Distance estimation · Experimental data | mean absolute distance error: 24 cm -> 20 cm (17%) | Helped |
| Add polarization information to wavefront and Mueller inputs (ablation) | Proposed normal reconstruction · Synthetic data | mean angular error: 9.47 -> 8.71 | Helped |
| Add the Mueller matrix to wavefront and polarization inputs (ablation) | Proposed normal reconstruction · Synthetic data | mean angular error: 9.73 -> 8.71 | Helped |
| Use the full raw wavefront window instead of window size 1 (ablation) | Proposed normal reconstruction · Synthetic data | mean angular error: 10.64 -> 8.71 | Helped |
Surprising findings
- Classical shape-from-polarization (DoP-based) performs very poorly on outdoor scenes, with mean angular error of 49.82° on synthetic data.
Limitations
- Sequential rotating-ellipsometry acquisition takes about 5 minutes per frame, so not real-time.
- Real-world evaluation uses only 60 frames and noisier ground truth from a Velodyne reference.
- Low degree-of-polarization regions remain challenging for prior methods and rely on the learned prior.
Experimental setup
Datasets: CARLA-based synthetic polarimetric wavefront dataset PolLidar real-world long-range dataset
Benchmarks: PolLidar synthetic dataset PolLidar real-world dataset
Metrics: mean angular error median angular error RMSE accuracy at 3/5/10 degrees mean absolute distance error
Baselines: SfP-DoP Baek et al. [6] PCA normal estimation argmax peak-finding ToF
Builds on: Baek et al. temporal-polarimetric ToF model shape from polarization CARLA lidar simulator TransUNet
Closest papers by meaning
- SPIDeRS: Structured Polarization for Invisible Depth and Reflectance Sensing (CVPR 2024 · similarity 0.84)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.83)
- Blurred LiDAR for Sharper 3D: Robust Handheld 3D Scanning with Diffuse LiDAR and RGB (CVPR 2025 · similarity 0.83)
- Snapshot Lidar: Fourier Embedding of Amplitude and Phase for Single-Image Depth Reconstruction (CVPR 2024 · similarity 0.83)
- SfPUEL: Shape from Polarization under Unknown Environment Light (NeurIPS 2024 · similarity 0.83)
- LiSu: A Dataset and Method for LiDAR Surface Normal Estimation (CVPR 2025 · similarity 0.83)
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.