RAPID: Reconstruction Accelerated by Polarization-Informed Depth for 2D Gaussian Splatting
Organizations: Northeast Electric Power University
Abstract
Efficient and accurate 3D surface reconstruction from multi-view images remains a fundamental challenge in Gaussian splatting. Many existing methods rely on sparse geometric initialization and iterative densification, resulting in substantial optimization overhead and difficulties in balancing rendering quality and geometric accuracy, particularly in reflective and weakly textured regions. To address these limitations, we propose RAPID, a geometry-first framework that uses multiview polarization cues to recover dense surface priors. We integrate surface normals into depth maps using sparse geometric anchors. These depth maps are then fused across views. Curvature-adaptive sampling and oriented Gaussian initialization then support efficient optimization. Experiments on the PISR dataset demonstrate an approximately speedup over PhyGaP, which is a state-of-the-art Gaussian-based reconstruction method. RAPID reduces the average optimization time from approximately 21 minutes (PhyGaP) to 4 minutes while maintaining competitive novel-view synthesis quality and mesh reconstruction accuracy. These results demonstrate the effectiveness of RAPID in accelerating Gaussian optimization from polarimetric multi-view inputs.
Figures & tables
| Method | B. Loong | R. Loong | L. Rabbit | S. Rabbit | Time | |||||
| CD | F-Score | CD | F-Score | CD | F-Score | CD | F-Score | |||
| Polarimetric Implicit | PANDORA [ 9 ] | 2.49 | 39.34% | 2.87 | 44.25% | 1.50 | 55.27% | 1.53 | 60.79% | 12h+ |
| NeRSP [ 16 ] | 2.42 | 41.35% | 2.38 | 47.74% | 1.34 | 68.66% | 0.86 | 77.69% | 12h+ | |
| PISR [ 5 ] | 1.19 | 54.50% | 1.64 | 72.88% | 1.36 | 75.92% | 0.89 | 75.55% | 24.5m | |
| Explicit | SuGaR [ 14 ] | 2.01 | 34.85% | 2.83 | 28.65% | 1.46 | 46.55% | 1.78 | 41.42% | 43.0m |
| 2DGS [ 19 ] | 2.46 | 26.54% | 2.41 | 23.70% | 1.43 | 40.43% | 2.41 | 34.03% | 25.2m | |
| Scene | Metric | SuGaR [ 14 ] | 2DGS [ 19 ] | GOF [ 43 ] | MILo [ 13 ] | PhyGaP [ 38 ] | Pol2DGS (Ours) |
|---|---|---|---|---|---|---|---|
| B. Loong | PSNR | 19.78 | 20.03 | 19.82 | 17.56 | 21.04 | 20.50 |
| SSIM | 0.508 | 0.628 | 0.573 | 0.441 | 0.702 | 0.635 | |
| R. Loong | PSNR | 21.68 | 22.75 | 22.15 | 18.25 | 23.76 | 23.59 |
| SSIM | 0.655 | 0.760 | 0.708 | 0.625 | 0.811 | 0.775 | |
| L. Rabbit | PSNR | 17.79 | 24.30 | 22.71 | 16.12 | 26.36 | 26.45 |
| SSIM | 0.759 | 0.882 | 0.841 | 0.717 | 0.920 | 0.873 |
| 2DGS [ 19 ] | Polarimetric Surface Priors | Insert | Split | Merge | Prune | CD | PSNR | |
|---|---|---|---|---|---|---|---|---|
| (a) | 2.46 | 20.03 | ||||||
| (b) | 2.09 | 20.51 | ||||||
| (c) | 2.01 | 20.37 | ||||||
| (d) | 2.14 | 20.48 | ||||||
| (e) | 2.00 | 20.37 | ||||||
| (f) | 2.10 | 20.47 |
| B.Loong | R.Loong | L.Rabbit | S.Rabbit | |
|---|---|---|---|---|
| w/ Pol. Normal Constraint | ||||
| w/o Pol. Normal Constraint | ||||
| Direct Depth Fusion |
| Scene | Dense Pol. Init. | Sparse COLMAP Init. [ 33 ] | Sampled Pol. Init. | |
|---|---|---|---|---|
| B. Loong | CD | 2.04 | 2.19 | 1.95 |
| Time | 349s | 291s | 267s | |
| R. Loong | CD | 2.06 | 2.29 | 1.94 |
| Time | 331s | 277s | 269s | |
| L. Rabbit | CD | 1.61 | 1.91 | 1.44 |
| Time | 338s | 277s | 253s |
3D Reconstruction
3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.
See all 1,025 papers