Research map / 3D vision / Depth and stereo
360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries
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 360Loc, a 360-degree image localization dataset with lidar-refined ground truth poses and cross-device pinhole and fisheye queries generated via virtual cameras, benchmarking retrieval, local feature matching, and absolute pose regression.
Problem
Visual localization with 360-degree cameras as reference databases is under-explored, and existing datasets lack 360 reference/query images and cross-device (pinhole, fisheye, 360) evaluation in ambiguous scenes.
Approach
Generate lower-FoV pinhole and fisheye query images from 360-degree images via virtual cameras (remap or rectify), enabling fair cross-device comparison and data augmentation for retrieval and absolute pose regressors.
Key results
- MS-Transformer with virtual camera augmentation (MS-Tvc2) reduces translation error by up to 79% and rotation error by up to 72% compared to MS-T.
- VC2 rectification raises pinhole image retrieval R@1 with NetVLAD from 0.23 to 0.50, and 360 queries reach R@1 of 0.79 with NetVLAD.
- 360 queries achieve the best local matching accuracy in most settings, especially at night, while pinhole queries with VC2 approach 360 query performance at high and medium accuracy levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply VC2 rectification to query/reference images for image retrieval (ablation) | NetVLAD image retrieval · 360Loc, pinhole, fisheye1, fisheye2, and fisheye3 queries | R@1: 0.23 -> 0.5; 0.42 -> 0.73; 0.45 -> 0.74; 0.57 -> 0.77 | Helped |
| Apply VC1 remapping of queries to the 360 domain for image retrieval (ablation) | NetVLAD image retrieval · 360Loc, pinhole queries | R@1: 0.23 -> 0.24 | Helped |
| Apply VC1 remapping of queries to the 360 domain for image retrieval (ablation) | NetVLAD image retrieval · 360Loc, fisheye1, fisheye2, and fisheye3 queries | R@1: 0.42 -> 0.51; 0.45 -> 0.54; 0.57 -> 0.63 | Helped |
| Apply VC1 or VC2 to cross-device local matching localization (ablation) | Feature-matching localization · 360Loc, pinhole and fisheye queries | High accuracy (0.25m, 2°) | Helped |
| Train APRs with virtual-camera images (VC2) in addition to 360 images (ablation) | PN and MS-T · 360Loc, cross-device queries | Translation error: up to 79% | Helped |
| Train APRs with virtual-camera images (VC2) in addition to 360 images (ablation) | PN and MS-T · 360Loc, cross-device queries | Rotation error: up to 72% | Helped |
| Use 360° queries rather than narrower-FoV queries | Visual localization methods · 360Loc, challenging scenes | Localization accuracy | Helped |
Surprising findings
- Pinhole queries with VC2 approach 360 queries at medium accuracy, but pinhole local matching is better than fisheye for high and medium levels despite worse image retrieval.
- Adding virtual camera images to APR training reduces 360-query error even though the number of 360 training images is unchanged.
Limitations
- Only four campus scenes, mostly limited in scale and location diversity.
- Local matching uses 360 images cropped to 1228x614 due to computation cost.
- APRs struggle to extrapolate beyond the training set.
Experimental setup
Datasets: 360Loc
Benchmarks: 360Loc
Metrics: recall@k precision@k localization accuracy at (0.25m,2°), (0.5m,5°), (5m,10°) median translation error (m) median rotation error (°)
Baselines: NetVLAD CosPlace OpenIBL AP-GeM SuperPoint+SuperGlue DISK+LightGlue SuperPoint+LightGlue PoseNet MS-Transformer
Builds on: HLoc PoseNet MS-Transformer SuperGlue LightGlue
Closest papers by meaning
- CroCoDL: Cross-device Collaborative Dataset for Localization (CVPR 2025 · similarity 0.87)
- GPVK-VL: Geometry-Preserving Virtual Keyframes for Visual Localization under Large Viewpoint Changes (CVPR 2025 · similarity 0.87)
- Enhancing Visual Localization with Cross-Domain Image Generation (ICML 2025 · similarity 0.86)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.85)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.85)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · 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.