Research map / 3D vision / Depth and stereo
Fully Geometric Panoramic 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.
Fully geometric panoramic localization that matches 2D lines from a panorama against a 3D line map using line distance and line-intersection point distance functions, with no visual descriptors or training.
Problem
Visual localization against 3D maps typically relies on visual descriptors, which are costly to store, raise privacy concerns, and fail under illumination or scene changes; line-only localization has been ambiguous and less accurate.
Approach
Uses line intersections labeled by principal direction clusters as point distance functions, decouples rotation and translation by caching canonicalized 3D distance functions and interpolating 2D distance functions, and refines poses by aligning intersection points and line directions, all without learned or hand-crafted descriptors.
Key results
- On the large-scale OmniScenes/Stanford splits, the method reaches 0.76/0.61/0.68/0.64 accuracy at (0.1m, 5°) with search time 0.02 s and map size 2.56 GB, versus LDL at 0.54 search and 129.75 GB.
- Pose refinement under the Extreme OmniScenes split achieves 0.67 accuracy at (0.1m, 5°), outperforming other geometric refiners such as GoMatch (0.61) and Gao et al. (0.21).
- On Structured3D floorplan localization, the method achieves 0.95 accuracy at (0.1m, 5°), above LaLaLoc (0.91) and Laser2D (0.79).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add point distance functions to line distance functions (ablation) | Pose search · Stanford 2D-3D-S Area 1 | accuracy (0.1m, 5°): 0.57 -> 0.61 | Helped |
| decouple rotation and translation for 3D distance functions (ablation) | Pose search · Stanford 2D-3D-S Area 1 | runtime (s), GPU: 45.049 -> 2.890 | Helped |
| decouple rotation and translation for 3D distance functions (ablation) | Pose search · Stanford 2D-3D-S Area 1 | accuracy (0.1m, 5°): 0.62 -> 0.61 | No clear effect |
| enable canonicalization of 2D distance functions (ablation) | Pose search · Stanford 2D-3D-S Area 1 | runtime (s), GPU: 0.396 -> 0.007 | Helped |
| add interpolation to canonicalized 2D distance functions (ablation) | Pose search · Stanford 2D-3D-S Area 1 | runtime (s), GPU: 0.007 -> 0.004 | Helped |
| add rotation refinement after translation refinement (ablation) | Pose refinement · OmniScenes Room 2 | rotation error: 1.56 -> 1.05 | Helped |
| add intersection cluster-guided matching (ablation) | Pose refinement · OmniScenes Room 2 | accuracy (0.1m, 5°): 0.63 -> 0.77 | Helped |
| use line intersections and their geometric matching instead of visual descriptors | Ours · OmniScenes and Stanford 2D-3D-S | accuracy (0.1m, 5°): 0.67 vs 0.79 | Mixed |
| localize using line geometry under lighting changes | Ours · OmniScenes, varying lighting conditions | accuracy (0.1m, 5°): 0.77, 0.74, 0.74, 0.77 | No clear effect |
Surprising findings
- The geometric-only pipeline is competitive with point-based refiners that use visual descriptors (e.g., 0.78 for SuperGlue at 0.1m/5°), despite using only lines.
- Under illumination changes the geometric method shows a much smaller accuracy range (0.05) than descriptor-based methods such as Line Transformer (0.22).
Limitations
- Relies on line extraction working; performance depends on line availability in the scene.
- Experiments use room-level or floorplan maps; the method's behavior in highly unstructured scenes is not evaluated.
Experimental setup
Datasets: Stanford 2D-3D-S OmniScenes Structured3D
Benchmarks: OmniScenes Stanford 2D-3D-S Structured3D
Metrics: localization accuracy at (0.1m, 5°) accuracy at (0.2m, 10°) accuracy at (0.3m, 15°) recall curves search runtime refinement runtime map size translation error rotation error
Baselines: LDL SFRS CosPlace PICCOLO CPO Chamfer LightGlue SuperGlue GlueStick GoMatch LoFTR Line Transformer
Builds on: LDL ICP Kabsch algorithm
Closest papers by meaning
- Unified Panoramic Geometry Estimation via Multi-View Foundation Models (NeurIPS 2026 · similarity 0.86)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.86)
- GPVK-VL: Geometry-Preserving Virtual Keyframes for Visual Localization under Large Viewpoint Changes (CVPR 2025 · similarity 0.86)
- Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images (CVPR 2025 · similarity 0.85)
- Learning to Produce Semi-dense Correspondences for Visual Localization (CVPR 2024 · similarity 0.85)
- Geometric Exploitation for Indoor Panoramic Semantic Segmentation (NeurIPS 2024 · similarity 0.84)
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.