Research map / 3D vision / Depth and stereo
Relative Pose Estimation through Affine Corrections of Monocular Depth Priors
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.
Proposes three minimal/over-determined solvers that estimate relative pose from 2D correspondences and monocular depth priors while modeling per-image shift and a relative scale, combined with classic point-based solvers in a hybrid LO-MSAC pipeline.
Problem
Relative pose estimation from feature correspondences is limited by noise and ambiguity in monocular depth priors, which are affine-invariant (scale and shift unknown), so prior methods assuming only scale ambiguity fail.
Approach
Model independent scale and shift of both monocular depth maps explicitly and solve for them jointly with relative pose via pairwise 3D distance consistency, yielding calibrated (3-point), shared-focal (4-point), and two-focal (4-point) solvers that are fused with point-based solvers in a hybrid LO-MSAC.
Key results
- Calibrated solver on ScanNet-1500 with SP+LG and MoGe priors reaches pose AUC@5 of 23.36 vs 21.55 for PoseLib 5-pt.
- Shared-focal solver with RoMa matches and DA-met. priors gives AUC@5 of 29.81 vs 27.17 for PoseLib 6-pt on ScanNet-1500 (median focal error 3.71% vs 4.49%).
- Two-focal solver on 2D-3D-S pairs with SP+LG reduces median rotation error to 5.66 from 16.83 (PoseLib 7-pt) and median translation error to 9.26 from 21.57.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use affine-corrected depth priors in the calibrated hybrid estimator | Relative pose estimation · ScanNet-1500 calibrated, RoMa | pose error AUC @5°: 33.40 -> 34.26 | Helped |
| Use affine-corrected depth priors in the calibrated hybrid estimator | Relative pose estimation · MegaDepth-1500 calibrated | pose error AUC @5°: 60.16 -> 63.48 | Helped |
| Use the shared-focal depth-aware hybrid solver | Relative pose estimation · ScanNet-1500 shared-focal, RoMa | pose error AUC @5°: 27.17 -> 29.81 | Helped |
| Use the two-focal depth-aware hybrid solver | Relative pose estimation · Stanford 2D-3D-S, MASt3R matches | pose error AUC @5°: 12.58 -> 18.05 | Helped |
| Use the two-focal depth-aware hybrid solver | Relative pose estimation · MegaDepth-1500 two-focal | pose error AUC @5°: 21.23 -> 25.70 | Helped |
| Add shift modeling (ablation) | Calibrated estimator with DA-met. priors · ScanNet-1500 | pose error AUC @5°: 19.96 -> 23.12 | Helped |
| Add hybrid estimation (ablation) | Calibrated estimator with DA-met. priors and shift modeling · ScanNet-1500 | pose error AUC @5°: 19.13 -> 23.12 | Helped |
| Add shift modeling (ablation) | Calibrated estimator with Marigold priors · ScanNet-1500 | pose error AUC @5°: 12.34 -> 20.68 | Helped |
| Add shift modeling (ablation) | Calibrated estimator with MoGe priors · MegaDepth-1500 | pose error AUC @5°: 51.32 -> 63.48 | Helped |
| Use hybrid solvers, local optimization, and scoring together (ablation) | Shared-focal LO-MSAC · ScanNet-1500, SP+SG | pose error AUC @5°: 13.85 -> 18.35 | Helped |
| Use hybrid scoring (ablation) | Shared-focal LO-MSAC with depth solver and LO · ScanNet-1500, SP+SG | pose error AUC @5°: 13.85 -> 15.42 | Helped |
| Use the depth-aware solver alongside point-based solvers (ablation) | Shared-focal hybrid LO-MSAC · ScanNet-1500, SP+SG | pose error AUC @5°: 16.60 -> 18.35 | Helped |
Surprising findings
- Modeling shift also improves results with metric depth models (Depth-Anything metric), not only affine-invariant ones.
- The 3p3d and 4p4d solvers from prior work fail to improve over point-based baselines because they ignore shift.
Limitations
- Solvers depend on quality of monocular depth priors; hybrid with point solvers needed when priors are unreliable.
- Outdoor performance is weaker with metric depth models (DA-met.) than with MoGe.
- Runtime increases due to depth-aware solvers and Ceres autodiff Jacobians.
Experimental setup
Datasets: ScanNet-1500 MegaDepth-1500 Stanford 2D-3D-S ETH3D
Benchmarks: ScanNet-1500 MegaDepth-1500 ETH3D Stanford 2D-3D-S
Metrics: median rotation error median translation direction error pose error AUC median relative focal error
Baselines: PoseLib PnP-RANSAC with monocular depth 2PT+D&5pt 3p3d 4p4d DUSt3R MASt3R
Builds on: LO-RANSAC hybrid RANSAC Larsson et al. Gröbner basis solver method 2PT+D solver
Closest papers by meaning
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.91)
- From Correspondences to Pose: Non-minimal Certifiably Optimal Relative Pose without Disambiguation (CVPR 2024 · similarity 0.89)
- DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.88)
- UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.88)
- FAR: Flexible Accurate and Robust 6DoF Relative Camera Pose Estimation (CVPR 2024 · similarity 0.88)
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.87)
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.