Research map / 3D vision / Gaussian splatting
Unlocking Accurate Geometry in 3DGS via Spatially Varying Affine Rectification
Tze Ho Elden Tse, Jizong Peng, Richard Chen, Angela Yao
Read the paperCheck your idea against 137,217 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.
Proposes spatially varying affine rectification of monocular depth priors to regularize 3D Gaussian Splatting for more accurate geometry.
Problem
3DGS produces photorealistic novel views but its geometry is inaccurate; existing depth-guided regularization uses monocular depth priors aligned with a single global scale-and-shift, which the authors argue is flawed.
Approach
Replaces the single global scale-and-shift alignment of monocular depth with spatially varying affine rectification, since modern monocular depth estimators are not consistent under one global affine transform.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add L_depth (global scale anchor) to Splatfacto (ablation) | Splatfacto · MuSHRoom coffee_room | Abs Rel: 0.059 -> 0.043 | Improved |
| add L_grad alone to Splatfacto (ablation) | Splatfacto · MuSHRoom coffee_room | F-score: 0.728 -> 0.673 | Hurt |
| add L_grad to L_normal + L_depth (ablation) | Full method · MuSHRoom coffee_room | F-score: 0.813 -> 0.855 | Improved |
| add L_normal to L_grad + L_depth (ablation) | Full method · MuSHRoom coffee_room | F-score: 0.850 -> 0.855 | No clear effect |
| add all geometric regularizers to Splatfacto (ablation) | Splatfacto · MuSHRoom coffee_room | PSNR: 25.293 -> 24.947 | Hurt |
| partition scene into K=5 local affine regions instead of one global affine (ablation) | Depth-Anything-V2 alignment · MuSHRoom | inlier ratio (0.02 threshold): 58.9% -> 92.7% | Improved |
| rectify monocular depth with spatially varying affine fields instead of global alignment | DN-Splatter (Monodepth) · MuSHRoom left split | Abs Rel: 0.3542 -> 0.0551 | Improved |
| rectify monocular depth instead of global alignment (mesh) | DN-Splatter (Monodepth) · MuSHRoom mesh | F-score: .2765 -> .7097 | Improved |
| rectified monocular prior vs sparse-view baselines | 3DGS · NeRF-LLFF mean, 2-view | PSNR: 17.31 -> 18.79 (vs DRGS) | Improved |
| rectified monocular prior on ScanNet++ meshes | Splatfacto · ScanNet++ mesh | F-score: .3207 vs SuGaR .4367 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DRGS — DR-GS: Physically-Based Deformable and Relightable 2D Gaussians (ECCV 2026) · better on 5 of 5 shared results
- DropGaussian — DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Surprising findings
- Modern monocular depth estimates are not consistent with a single global scale-and-shift alignment, the core assumption of existing depth-guided 3DGS methods.
Experimental setup
Builds on: 3D Gaussian Splatting
Closest papers by meaning
- NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.90)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.90)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.90)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
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.
