Research map / 3D vision / Depth and stereo
HyperPose: Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset
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 HyperPose, which adds a hypernetwork that generates regression-head weights per query image for single- and multi-scene absolute pose regressors, and releases the Extended Cambridge Landmarks (ECL) dataset with synthetic appearance variants.
Problem
Absolute camera pose regression (APR) loses accuracy under appearance variation (lighting, season, viewpoint) between training and test images, and single-network APRs generalize poorly across scenes.
Approach
A hypernetwork takes the CNN feature of the query image and outputs the weights of the position and orientation regression heads, so the head is adapted per input; the same idea is applied to a multi-scene transformer-based APR (MS-HyperPose).
Key results
- Baseline APR with hypernetwork reduces Cambridge average median error from 1.13m/3.84° to 1.08m/3.57°, and AtLoc from 1.69m/4.39° to 1.40m/4.18°.
- MS-HyperPose reaches the best multi-scene Cambridge average of 1.27m/2.41° versus MS-Trans at 1.28m/2.73°.
- On 7 Scenes, the hypernetwork baseline averages 0.19m/7.48°, and MS-HyperPose averages 0.18m/7.00° (reported as 7.12° in text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a hypernetwork to Baseline APR (ablation) | Baseline APR · Cambridge Landmarks (Avg.) | median position error (m): 1.13 -> 1.08 | Helped |
| Add a hypernetwork to Baseline APR (ablation) | Baseline APR · Cambridge Landmarks (Avg.) | median orientation error (degrees): 3.84 -> 3.57 | Helped |
| Add a hypernetwork to AtLoc (ablation) | AtLoc · Cambridge Landmarks (Avg.) | median position error (m): 1.69 -> 1.40 | Helped |
| Add a hypernetwork to AtLoc (ablation) | AtLoc · Cambridge Landmarks (Avg.) | median orientation error (degrees): 4.39 -> 4.18 | Helped |
| Add a hypernetwork to Baseline APR (ablation) | Baseline APR · 7 Scenes | median orientation error (degrees) | Mixed |
| Add hypernetworks to MS-Trans (ablation) | MS-Trans · Cambridge Landmarks (Avg.) | median orientation error (degrees): 2.73 -> 2.41 | Helped |
| Add hypernetworks to MS-Trans (ablation) | MS-Trans · 7 Scenes (Avg.) | median position error (m): 0.18 -> 0.18 | No clear effect |
| Add hypernetworks to MS-Trans (ablation) | MS-Trans · 7 Scenes (Avg.) | median orientation error (degrees): 7.28 -> 7.00 | Helped |
| Add a hypernetwork to AtLoc (ablation) | AtLoc · Oxford RobotCar LOOP1 | median position error (m): 5.68 -> 4.91 | Helped |
| Add a hypernetwork to AtLoc (ablation) | AtLoc · Oxford RobotCar FULL1 | median position error (m): 11.10 -> 7.86 | Helped |
| Add a hypernetwork to AtLoc (ablation) | AtLoc · Oxford RobotCar FULL2 | median position error (m): 12.2 -> 8.38 | Helped |
| Add a hypernetwork to Baseline APR (ablation) | Baseline APR · Extended Cambridge Landmarks (ECL) | pose localization error | Helped |
Surprising findings
- MS-HyperPose model size is 571MB versus 71.1MB for the original MS-Transformer, a large storage overhead.
Limitations
- Substantially larger model size for multi-scene variant (571MB vs 71.1MB)
- Improvements over baselines are modest on some scenes, and some metrics are not best (e.g., MS-HyperPose worse on some Cambridge orientation values)
- Single-scene results reported without the two additional fine-tuning stages
Experimental setup
Datasets: Cambridge Landmarks 7 Scenes Oxford RobotCar
Benchmarks: Cambridge Landmarks 7 Scenes Oxford RobotCar Extended Cambridge Landmarks (ECL)
Metrics: median position error (m) median orientation error (degrees) mean position error mean orientation error inference time (ms) model size (MB)
Baselines: AtLoc PoseNet MSPN MS-Trans MapNet DSAC* IRPNet GeoPoseNet
Builds on: PoseNet AtLoc MS-Transformer Multi-Scene PoseNet Kendall uncertainty multi-task loss InstructPix2Pix
Closest papers by meaning
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.87)
- Map-Relative Pose Regression for Visual Re-Localization (CVPR 2024 · similarity 0.87)
- GLACE: Global Local Accelerated Coordinate Encoding (CVPR 2024 · similarity 0.85)
- ConDo: Continual Domain Expansion for Absolute Pose Regression (AAAI 2025 · similarity 0.85)
- R-SCoRe: Revisiting Scene Coordinate Regression for Robust Large-Scale Visual Localization (CVPR 2025 · similarity 0.85)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · 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.