AffineGS: Unlocking Accurate Geometry in 3DGS via Spatially Varying Affine Rectification
Organizations: 1dConstruct Robotics · 2National University of Singapore
Abstract
3D Gaussian Splatting (3DGS) has become the leading approach for photorealistic novel view synthesis, yet its geometric accuracy lags behind its visual quality. Existing depth-guided methods attempt to resolve this by regularizing the optimization with monocular depth priors, typically using a single global scale-and-shift alignment. Crucially, we discover a fundamental flaw in this assumption: modern monocular depth networks systematically exhibit region-dependent, piecewise-linear distortions. This discovery invalidates standard global alignment strategies, which warp scene geometry by enforcing a single transform. To address this, we introduce a spatially varying affine rectification model that corrects monocular disparity into consistent metric depth using regionally smooth, edge-aware fields. We further derive a normal-aware regularizer that uses spatial gradients to couple the rectified disparity with 3DGS depth and normals. Across diverse experimental settings, our sensor-free method shows strong photometric quality of Gaussian Splatting while substantially improving geometric accuracy and outperforming state-of-the art geometry-focused methods.
Figures & tables
| MuSHRoom | ScanNet++ | |||
|---|---|---|---|---|
| Alignment Strategy | Inlier Ratio ( ) | RMSE ( ) | Inlier Ratio ( ) | RMSE ( ) |
| Global Affine ( ) | 58.9% | 0.074 | 58.6% | 0.088 |
| Piecewise Local ( ) | 84.9% | 0.046 | 84.6% | 0.056 |
| Piecewise Local ( ) | 92.7% | 0.030 | 91.8% | 0.039 |
| Method | Photometry | Geometry | |||
| PSNR | LPIPS | Abs Rel | Accuracy | F-Score | |
| Splatfacto (Baseline) | 25.293 | 0.086 | 0.059 | 0.043 | 0.728 |
| + | 24.322 | 0.123 | 0.085 | 0.054 | 0.673 |
| + | 25.256 | 0.085 | 0.064 | 0.042 | 0.728 |
| + | 24.561 | 0.098 | 0.043 | 0.030 | 0.798 |
| + + | 24.553 | 0.116 | 0.075 | 0.051 | 0.707 |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Strict Threshold (0.005) | Moderate Threshold (0.02) | Forgiving Threshold (0.05) | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Alignment Strategy | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) |
| Global Affine ( ) | 69.4% | 0.518 | 0.252 | 77.0% | 0.517 | 0.253 | 92.2% | 0.517 | 0.254 |
| Piecewise Local ( ) | 88.7% | 0.515 | 0.271 | 92.6% | 0.513 | 0.274 | 98.1% | 0.373 | 0.452 |
| Piecewise Local ( ) | 92.1% | 0.502 | 0.289 | 94.8% | 0.463 | 0.329 | 98.8% | 0.253 | 0.635 |
| Piecewise Local ( ) | 94.0% | 0.454 | 0.338 | 96.1% | 0.394 | 0.414 | 99.1% | 0.181 | 0.759 |
| Piecewise Local ( ) | 95.3% | 0.406 | 0.399 | 97.0% | 0.326 | 0.509 | 99.3% | 0.137 | 0.839 |
| Strict Threshold (0.005) | Moderate Threshold (0.02) | Forgiving Threshold (0.05) | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Alignment Strategy | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) |
| Global Affine ( ) | 44.1% | 0.042 | 0.982 | 77.4% | 0.040 | 0.981 | 92.1% | 0.039 | 0.982 |
| Piecewise Local ( ) | 70.5% | 0.032 | 0.986 | 93.7% | 0.024 | 0.991 | 98.6% | 0.020 | 0.994 |
| Piecewise Local ( ) | 76.9% | 0.029 | 0.988 | 96.0% | 0.021 | 0.993 | 99.2% | 0.017 | 0.996 |
| Piecewise Local ( ) | 81.4% | 0.026 | 0.989 | 97.1% | 0.018 | 0.995 | 99.5% | 0.015 | 0.997 |
| Piecewise Local ( ) | 84.7% | 0.024 | 0.991 | 97.9% | 0.015 | 0.996 | 99.7% | 0.013 | 0.998 |
| Strict Threshold (0.005) | Moderate Threshold (0.02) | Forgiving Threshold (0.05) | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Alignment Strategy | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) |
| Global Affine ( ) | 25.6% | 0.086 | 0.967 | 58.9% | 0.074 | 0.967 | 82.4% | 0.071 | 0.967 |
| Piecewise Local ( ) | 50.6% | 0.060 | 0.974 | 84.9% | 0.046 | 0.980 | 96.3% | 0.033 | 0.990 |
| Piecewise Local ( ) | 58.4% | 0.055 | 0.976 | 89.7% | 0.037 | 0.985 | 98.0% | 0.026 | 0.994 |
| Piecewise Local ( ) | 64.4% | 0.050 | 0.978 | 92.7% | 0.030 | 0.990 | 98.8% | 0.022 | 0.996 |
| Piecewise Local ( ) | 69.2% | 0.047 | 0.980 | 94.6% | 0.025 | 0.994 | 99.2% | 0.019 | 0.997 |
| Strict Threshold (0.005) | Moderate Threshold (0.02) | Forgiving Threshold (0.05) | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Alignment Strategy | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) | Inlier % ( ) | RMSE ( ) | Corr. ( ) |
| Global Affine ( ) | 26.9% | 0.092 | 0.952 | 58.6% | 0.088 | 0.952 | 81.6% | 0.084 | 0.952 |
| Piecewise Local ( ) | 51.8% | 0.068 | 0.967 | 84.6% | 0.056 | 0.974 | 95.2% | 0.040 | 0.985 |
| Piecewise Local ( ) | 59.2% | 0.065 | 0.969 | 89.1% | 0.047 | 0.980 | 97.1% | 0.031 | 0.991 |
| Piecewise Local ( ) | 64.8% | 0.062 | 0.971 | 91.8% | 0.039 | 0.985 | 98.1% | 0.025 | 0.994 |
| Piecewise Local ( ) | 69.3% | 0.058 | 0.974 | 93.7% | 0.032 | 0.989 | 98.8% | 0.022 | 0.995 |
| Dataset | Scene | Radius (px) | ||||
|---|---|---|---|---|---|---|
| MuSHRoom | coffee_room | 600 | 0.08 | 0.1 | 0.08 | 0.4 |
| computer | 600 | 0.5 | 0.1 | 0.08 | 0.4 | |
| honka | 600 | 0.1 | 0.1 | 0.08 | 0.4 | |
| kokko | 550 | 0.4 | 0.05 | 0.04 | 0.2 | |
| sauna | 550 | 0.3 | 0.025 | 0.02 | 0.1 | |
| vr_room | 650 | 0.2 | 0.1 | 0.08 | 0.4 |
| Metric | Mathematical Formulation |
|---|---|
| Accuracy | |
| Completion | |
| Chamfer- | |
| Normal Accuracy | s.t. |
| Normal Completion | s.t. |
| Normal Consistency |
| Method | Photometry | Depth Estimation | Mesh Reconstruction | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PSNR | SSIM | LPIPS | Abs Rel | Sq Rel | RMSE | Acc. | Comp. | Chamfer- | Norm. Cons. | F-score | ||
| Splatfacto | 25.293 | 0.869 | 0.086 | 0.060 | 0.021 | 0.196 | 0.937 | 0.043 | 0.039 | 0.041 | 0.779 | 0.728 |
| + | 24.323 | 0.848 | 0.123 | 0.085 | 0.045 | 0.262 | 0.889 | 0.055 | 0.051 | 0.053 | 0.857 | 0.673 |
| + | 25.257 | 0.867 | 0.085 | 0.064 | 0.022 | 0.205 | 0.926 | 0.043 | 0.039 | 0.041 | 0.797 | 0.728 |
| + | 24.561 | 0.863 | 0.098 | 0.043 | 0.011 | 0.147 | 0.984 | 0.031 | 0.039 | 0.035 | 0.793 | 0.799 |
| + + | 24.553 | 0.851 | 0.117 | 0.075 | 0.034 | 0.238 | 0.902 | 0.051 | 0.047 | 0.049 | 0.868 | 0.707 |
| Depth Prior | Scene | Photometry | Depth Estimation | Mesh Reconstruction | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PSNR | SSIM | LPIPS | Abs Rel | Sq Rel | RMSE | Acc. | Comp. | Chamfer- | Norm. Cons. | F-score | |||
| Depth-Anything-V2 | coffee_room | 24.947 | 0.858 | 0.103 | 0.040 | 0.010 | 0.139 | 0.985 | 0.026 | 0.028 | 0.027 | 0.891 | 0.855 |
| computer | 24.968 | 0.851 | 0.129 | 0.070 | 0.026 | 0.254 | 0.955 | 0.065 | 0.058 | 0.061 | 0.884 | 0.603 | |
| honka | 23.142 | 0.821 | 0.140 | 0.050 | 0.017 | 0.175 | 0.973 | 0.047 | 0.034 | 0.041 | 0.885 | 0.799 | |
| kokko | 22.444 | 0.831 | 0.163 | 0.064 | 0.025 | 0.220 | 0.933 | 0.051 | 0.042 | 0.047 | 0.798 | 0.705 | |
| sauna | 24.443 | 0.830 | 0.187 | 0.060 | 0.015 | 0.167 | 0.978 | 0.051 | 0.044 | 0.047 | 0.874 | 0.687 | |
| Scene | Photometry (Left / Right) | Depth Estimation (Left / Right) | Mesh Reconstruction | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PSNR | SSIM | LPIPS | Abs Rel | Sq Rel | RMSE | Acc. | Comp. | Chamfer- | Norm. Cons. | F-score | ||
| coffee_room | 22.71 / 19.74 | 0.788 / 0.727 | 0.187 / 0.288 | 0.088 / 0.086 | 0.033 / 0.033 | 0.245 / 0.232 | 0.862 / 0.903 | 0.076 | 0.084 | 0.080 | 0.792 | 0.516 |
| computer | 13.60 / 13.27 | 0.637 / 0.602 | 0.495 / 0.506 | 0.744 / 0.635 | 1.791 / 1.875 | 2.273 / 2.541 | 0.100 / 0.242 | 0.276 | 0.666 | 0.471 | 0.641 | 0.065 |
| honka | 21.38 / 19.40 | 0.755 / 0.705 | 0.218 / 0.272 | 0.101 / 0.091 | 0.047 / 0.028 | 0.275 / 0.244 | 0.877 / 0.926 | 0.124 | 0.109 | 0.117 | 0.811 | 0.409 |
| kokko | 12.81 / 11.38 | 0.617 / 0.592 | 0.537 / 0.585 | 0.823 / 0.921 | 1.649 / 2.256 | 1.958 / 2.532 | 0.046 / 0.009 | 0.445 | 0.689 | 0.567 | 0.524 | 0.027 |
| sauna | 20.95 / 18.52 | 0.745 / 0.718 | 0.329 / 0.373 | 0.284 / 0.356 | 0.319 / 0.443 | 0.596 / 0.708 | 0.619 / 0.545 | 0.136 | 0.240 | 0.188 | 0.724 | 0.316 |
3D Gaussian Splatting
3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.
See all 1,094 papers