RF3DGS: Reflection-Free 3D Gaussian Splatting with Self-Priors
Organizations: The Hong Kong University of Science and Technology · Huawei Research
Abstract
3D reflection removal is crucial in applications demanding high-fidelity visualization, such as museum exhibitions. In this work, we propose Reflection-Free 3D Gaussian Splatting (RF3DGS), a novel method for superior reflection removal. RF3DGS exclusively leverages a widely-adopted photometric prior for reflection removal: that the minimum pixel intensity across multi-view images corresponding to the same 3D point strongly correlates with reflection-free information. However, generalizing this principle to a 3D Gaussian Splatting within a multi-view context is non-trivial. To address this, RF3DGS first optimizes a vanilla 3DGS model using a geometric prior to estimate a coarse, reflection-free depth map for each view. Leveraging these depth maps and photometric prior, a 3D warping process aggregates reflection-free information from auxiliary views to construct a clean transmission pseudo-label for a given reference view. This pseudo-label provides a robust self-supervised signal for optimization. We further introduce an Adaptive Twin Rendering (ATR) loss to synergistically utilize this pseudo-label while preserving high-frequency details from the original, reflection-degraded images. By adeptly fusing these internal 3D self-priors with the ATR loss, RF3DGS demonstrates state-of-the-art reflection removal performance across both synthetic and real-world scenarios. Furthermore, we establish the first 3D reflection removal benchmark, based on the Unity rendering engine, for rigorous quantitative evaluation.
Figures & tables
| Configuration | Intensity Range | Total Cases |
| Transparency (%) | 45 | |
| Thickness (mm) | ||
| Reflection Strength |
| Method | PSNR | SSIM | LPIPS |
| Chugunov et al . [ 4 ] | 2.94 | 0.248 | 0.935 |
| Vanilla 3DGS [ 17 ] | 29.58 | 0.917 | 0.175 |
| DSRNet [ 14 ] + Vanilla 3DGS | 26.91 | 0.872 | 0.240 |
| NeRFReN [ 10 ] | 26.36 | 0.755 | 0.284 |
| Ours | 31.45 | 0.915 | 0.150 |
| Method | PSNR | SSIM | LPIPS |
| MVSNeRF [ 3 ] | 17.83 | 0.685 | 0.400 |
| NeRFReN [ 10 ] | 18.63 | 0.716 | 0.449 |
| TNeRF [ 54 ] | 22.75 | 0.841 | 0.205 |
| Vanilla 3DGS [ 17 ] | 21.73 | 0.835 | 0.238 |
| DSRNet [ 14 ] + Vanilla 3DGS | 22.03 | 0.838 | 0.221 |
| Flash-Splat [ 47 ] | N/A | N/A | N/A |
| Method | PSNR | SSIM | LPIPS |
| Ours ( w/o ) | 19.64 | 0.790 | 0.237 |
| Ours ( w/o ) | 30.45 | 0.888 | 0.176 |
| Ours (w/o ) | 31.00 | 0.908 | 0.155 |
| Ours (w/o ) | 31.11 | 0.909 | 0.155 |
| Ours | 31.25 | 0.915 | 0.150 |
| Ours (11 views) | 31.14 | 0.907 | 0.140 |
| Vanilla 3DGS | DSRNet + Vanilla 3DGS | NeRFReN | Ours | Ours (3D Warp) | |
| Time | 5.94 min | 7.85 min | 63.48 h | 60.74 min | 30 sec |
3D Gaussian Splatting
3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.
See all 1,093 papers