Research map / 3D vision / Human pose estimation
Scene-agnostic Pose Regression for Visual Localization
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.
Introduces Scene-agnostic Pose Regression (SPR), a 360° panoramic dataset (360SPR) of 200K panoramas across 270 Habitat scenes, and SPR-Mamba, a dual-branch Mamba model that regresses query poses relative to the first frame without retraining or retrieval databases.
Problem
Visual localization methods trade off generalization against database dependence and drift: absolute pose regression needs retraining for new scenes, relative pose regression needs a large retrieval database, and visual odometry accumulates drift over open trajectories.
Approach
Treat the first frame of a sequence as the coordinate origin and regress each query pose relative to that origin using all preceding frames, with a dual-branch Mamba model (local frame-to-frame branch plus global query-to-origin branch), so there is neither retrieval nor accumulated drift.
Key results
- On 360SPR unseen scenes, SPR-Mamba achieves 3.85m / 3.97° average median TE/RE, reducing error by over 23m/43° versus APR and over 7m/16° versus RPR.
- On 360Loc unseen scenes, SPR-Mamba achieves 1.94m / 1.44° average median TE/RE, outperforming RPR by over 7m/17°.
- Ablation on 360SPR shows the full model (local + global branches with auxiliary heads) improves unseen TE/RE from 10.33/10.68 (global branch only) to 3.85/3.97.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPR-Mamba instead of the strongest listed VO baseline, LEAP-VO | 360SPR unseen localization · 360SPR unseen | translation error (m, average median): 3.89±0.3 -> 3.85±0.3 | Helped |
| Use SPR-Mamba instead of the strongest listed baseline, XVO | 360SPR unseen localization · 360SPR unseen | rotation error (°, average median): 4.21±0.3 -> 3.97±0.4 | Helped |
| Use SPR-Mamba instead of the strongest listed VO baseline, DPVO | 360Loc unseen localization · 360Loc unseen | translation error (m, average median): 2.40±0.3 -> 1.94±0.3 | Helped |
| Use SPR-Mamba instead of the strongest listed VO baseline, DPVO | 360Loc unseen localization · 360Loc unseen | rotation error (°, average median): 1.84±0.2 -> 1.44±0.2 | Helped |
| add the global branch (ablation) | SPR-Mamba · 360SPR unseen | translation error (m): 10.33±1.3 -> 3.85±0.3 | Helped |
| add the global branch (ablation) | SPR-Mamba · 360SPR unseen | rotation error (°): 10.68±1.4 -> 3.97±0.4 | Helped |
| add the local branch and auxiliary regression heads (ablation) | SPR-Mamba without local branch and auxiliary heads · 360SPR unseen | translation error (m): 4.72±0.7 -> 3.85±0.3 | Helped |
| add the local branch and auxiliary regression heads (ablation) | SPR-Mamba without local branch and auxiliary heads · 360SPR unseen | rotation error (°): 5.11±0.5 -> 3.97±0.4 | Helped |
| add auxiliary regression heads (ablation) | SPR-Mamba with local and global branches · 360SPR unseen | translation error (m): 4.32±0.6 -> 3.85±0.3 | Helped |
| add auxiliary regression heads (ablation) | SPR-Mamba with local and global branches · 360SPR unseen | rotation error (°): 4.67±0.5 -> 3.97±0.4 | Helped |
| train on mixed sensor heights instead of one sensor height (ablation) | SPR-Mamba · 360SPR unseen, cross-sensor evaluation | translation error (m) | Helped |
Surprising findings
- On 360SPR seen scenes, SPR-Mamba performs on par with the best image-based APR method Marepo.
- VO-style models trained on 360Loc generalize poorly to the more diverse 360SPR data, showing single-sensor-height and few-scene datasets are limiting.
Limitations
- SPR poses are relative to the origin frame; no absolute pose information is available.
- Handling of image distortions in panoramas is not yet robust.
- Experiments are on simulated photorealistic panoramas with indoor scenes only.
Experimental setup
Datasets: 360SPR 360Loc
Benchmarks: 360SPR 360Loc
Metrics: average median translation error (m) average median rotation error (deg) average mean translation error (m) average mean rotation error (deg)
Baselines: PoseNet NeFeS Marepo VidLoc MapNet GL-Net NN-Net FAR PanoPose DPVO LEAP-VO XVO
Builds on: Mamba DINO PoseNet DPVO
Closest papers by meaning
- Map-Relative Pose Regression for Visual Re-Localization (CVPR 2024 · similarity 0.90)
- ConDo: Continual Domain Expansion for Absolute Pose Regression (AAAI 2025 · similarity 0.88)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.86)
- GLACE: Global Local Accelerated Coordinate Encoding (CVPR 2024 · similarity 0.85)
- Neural Refinement for Absolute Pose Regression with Feature Synthesis (CVPR 2024 · similarity 0.85)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.85)
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.