Research map / 3D vision / Depth and stereo
Revisiting Global Translation Estimation with Feature Tracks
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 HETA, a hybrid explicit global translation averaging framework that jointly estimates cameras and 3D points from re-estimated relative translations and selected feature tracks, using a cross-product objective with L1-then-L2 optimization.
Problem
Global translation estimation in global SfM is inaccurate in low-parallax and collinear-motion scenes when relying only on relative translations, while feature-track-based methods are sensitive to feature track outliers.
Approach
Uses an explicit formulation estimating camera positions and 3D points together with a cross-product-distance objective (better than scale-form), re-estimates relative translations via epipolar coplanarity weighted by parallax (sin α), and selects reliable feature tracks sorted by maximum parallax angle.
Key results
- On KITTI, HETA achieves the lowest camera position errors on most sequences after BA (e.g., sequence 00 median 3.0 m vs 7.4 m for LiGT-style/PGILP baseline, mean 7.7 m vs 14.6 m).
- On 1DSfM, HETA registers the most images for most datasets and 1DSfM (explicit baseline) fails to reconstruct most scenes.
- Cross-product objective outperforms scale-form objective on KITTI, and hybrid input outperforms pure relative translations or pure feature tracks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HETA instead of CReTA-BATA | global translation estimation · KITTI-00 | median camera position error (m): 14.4 -> 2.4 | Helped |
| Use hybrid inputs and explicit 3D points instead of relative translations only (ablation) | cross-product objective on KITTI-00 · KITTI-00 | median camera position error (m): 5.0 -> 3.0 | Helped |
| Use hybrid inputs and explicit 3D points instead of pure feature tracks and implicit 3D points (ablation) | KITTI-00 translation estimation · KITTI-00 | median camera position error (m): 7.4 -> 3.0 | Helped |
| Use cross-product-form rather than scale-form objective (ablation) | translation estimation on KITTI · KITTI | camera position error | Helped |
| Use hybrid inputs rather than only relative translations or only feature tracks (ablation) | global translation estimation · KITTI and 1DSfM | camera position error | Helped |
| Estimate 3D points explicitly rather than implicitly (ablation) | translation estimation · KITTI and 1DSfM | camera position error | Helped |
| Increase the number of selected feature tracks (ablation) | HETA · KITTI and 1DSfM | camera position error | Mixed |
| Use unnormalized epipolar-plane normal vectors (ablation) | relative translation re-estimation · KITTI-06, KITTI-09, 1DSfM-PIC, and 1DSfM-ROF | relative translation angle error | Helped |
| Filter feature matches with parallax angles below 1.5° (ablation) | relative translation re-estimation · KITTI-06, KITTI-09, 1DSfM-PIC, and 1DSfM-ROF | relative translation angle error | Helped |
| Optimize with L1 norm before unbiased L2 refinement | HETA · KITTI and 1DSfM | camera position error | Helped |
| Use more feature tracks (ablation) | HETA · KITTI and 1DSfM | camera position error | Hurt |
Surprising findings
- The explicit 1DSfM method fails to reconstruct most of the KITTI data despite using feature tracks.
- Explicit methods are comparable in efficiency to implicit methods, contrary to the common belief that they increase problem complexity.
Limitations
- Performance still degraded by feature match outliers, which the authors identify as an open challenge.
- Relies on global rotations from Chatterjee's method, whose errors degrade translation accuracy on unordered data (1DSfM-ROF).
Experimental setup
Datasets: KITTI odometry 1DSfM
Benchmarks: KITTI odometry 1DSfM
Metrics: median camera position error (m) mean camera position error (m) number of registered cameras
Baselines: LUD CReTA-BATA LiGT PGILP 1DSfM
Builds on: 1DSfM LUD LiGT PGILP CReTA
Closest papers by meaning
- Sensitivity in Translation Averaging (NeurIPS 2023 · similarity 0.87)
- Learning to Filter Outlier Edges in Global SfM (CVPR 2025 · similarity 0.85)
- Baseline Desensitizing in Translation Averaging (CVPR 2018 · similarity 0.85)
- FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation (CVPR 2024 · similarity 0.83)
- MMTIT-Bench: A Multilingual and Multi-Scenario Benchmark with Cognition-Perception-Reasoning Guided Text-Image Machine Translation (CVPR 2026 · similarity 0.82)
- GLACE: Global Local Accelerated Coordinate Encoding (CVPR 2024 · similarity 0.82)
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.