Research map / 3D vision / Human pose estimation
Cycle-Sync: Robust Global Camera Pose Estimation through Enhanced Cycle-Consistent Synchronization
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 Cycle-Sync, a global camera pose estimator whose location solver adapts message-passing least squares with a Welsch-type robust loss and cycle-consistent edge weighting, plus a subspace-recovery outlier filter, avoiding bundle adjustment.
Problem
Global camera pose estimation (rotations and locations) from noisy, heavily corrupted relative rotations and unit direction vectors, especially under cycle-consistent corruption, without relying on bundle adjustment.
Approach
Adapt MPLS to location estimation: a Welsch-type loss rho(x)=1-exp(-a|x|) optimized by IRLS whose edge weights blend residuals with cycle-inconsistency scores s_ij computed from distances estimated in previous iterations (with a T-AAB initialization), giving exact-recovery guarantees using cycle consistency alone.
Key results
- On ETH3D, Cycle-Sync reduces median location error by 60.9% relative to LUD (both with STE), 66.1% vs BATA, 89.8% vs ShapeFit and 90.0% vs FusedTA under a standardized pipeline.
- Full Cycle-Sync pipeline reports median location error below 0.05 on average across ETH3D, versus averages above 0.2 for other SfM pipelines (Theia, GLOMAP, LUD).
- On synthetic uniform corruption without noise, Cycle-Sync achieves exact recovery up to corruption probability q≤0.8, while methods other than ShapeFit fail beyond q≈0.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Cycle-Sync instead of LUD for location estimation (ablation) | ETH3D location estimation without STE · ETH3D | median location error: 66.0% | Helped |
| replace REAPER with STE and filter outlying directions (ablation) | Cycle-Sync pipeline · ETH3D | median location error: 70.5% | Helped |
| upgrade IRLS rotation averaging to MPLS (ablation) | LUD pipeline · ETH3D | median location error: 17.8% | Helped |
| upgrade MPLS rotation averaging to MPLS-cycle (ablation) | LUD pipeline · ETH3D | median location error: 9.0% | Helped |
| use Cycle-Sync location estimation instead of LUD | ETH3D location estimation with STE · ETH3D | median location error: 60.9% | Helped |
| use Cycle-Sync instead of BATA | ETH3D location estimation with STE · ETH3D | median location error: 66.1% | Helped |
| use Cycle-Sync instead of ShapeFit | ETH3D location estimation with STE · ETH3D | median location error: 89.8% | Helped |
| use Cycle-Sync instead of FusedTA | ETH3D location estimation with STE · ETH3D | median location error: 90.0% | Helped |
| use Cycle-Sync location estimation | synthetic UCM location recovery · UCM, σ=0 | exact-recovery corruption threshold: q ≤ 0.8 | Helped |
| use Cycle-Sync location estimation | synthetic cycle-consistent corruption setting · adversarial corruption | robustness to corruption rate: q < 0.5 | Helped |
| use Cycle-Sync for reconstruction | ETH3D sparse reconstruction · ETH3D | triangulated 3D points: 9 out of the 13 datasets | Mixed |
Surprising findings
- Cycle-Sync yields fewer triangulated 3D points than GLOMAP or Theia on 4 of 13 ETH3D datasets (e.g., relief, relief_2), and all three methods perform poorly on meadow and office.
Limitations
- Theoretical guarantees cover only the initialization phase, not convergence of the full nonconvex optimization.
- Relies on the presence of well-shaped 3-cycles and may degrade on sparse or structured graphs lacking them.
- Hyperparameters (reweighting schedule, robust loss parameters) are manually selected.
Experimental setup
Datasets: ETH3D
Benchmarks: ETH3D
Metrics: median translation error number of triangulated 3D points
Baselines: LUD BATA ShapeFit FusedTA Theia GLOMAP
Builds on: MPLS CEMP AAB ShapeFit LUD
Closest papers by meaning
- Estimation of Camera Locations in Highly Corrupted Scenarios: All About That Base, No Shape Trouble (CVPR 2018 · similarity 0.87)
- Tensor-Based Synchronization and the Low-Rankness of the Block Trifocal Tensor (NeurIPS 2024 · similarity 0.86)
- Learning to Filter Outlier Edges in Global SfM (CVPR 2025 · similarity 0.85)
- RUBIK: A Structured Benchmark for Image Matching across Geometric Challenges (CVPR 2025 · similarity 0.84)
- ProbPose: A Probabilistic Approach to 2D Human Pose Estimation (CVPR 2025 · similarity 0.83)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.83)
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.