Research map / 3D vision / Human pose estimation
Real-Time RGB-D Camera Pose Estimation in Novel Scenes Using a Relocalisation Cascade
Tommaso Cavallari, Stuart Golodetz, Nicholas A. Lord, Julien Valentin, Victor Adrian Prisacariu, Luigi Di Stefano, Philip H. S. Torr
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ICP pose refinement (ablation) | Ours (Default) · 7-Scenes (average) | frames within 5cm/5deg (%): 93.41% -> 94.34% | Improved |
| add model-based hypothesis ranking of last 16 RANSAC candidates (ablation) | Ours (Default) + ICP · 7-Scenes (average) | frames within 5cm/5deg (%): 94.34% -> 96.41% | Improved |
| add hypothesis ranking (ablation) | Ours (Default) + ICP · 7-Scenes Stairs | frames within 5cm/5deg (%): 80.00% -> 94.20% | Improved |
| add hypothesis ranking (cost) (ablation) | Ours (Default) + ICP · 7-Scenes frame time | frame time (ms): 133 -> 257 | Hurt |
| use relocalisation cascade F->S instead of single Default+Ranking | Ours (Fast) · 7-Scenes (average) | frames within 5cm/5deg (%): 90.83% -> 94.95% | Improved |
| use cascade F->S instead of Default+Ranking for speed | Ours (Default) + Ranking · 7-Scenes frame time | frame time (ms): 257 -> 52 | Improved |
| replace pre-trained forest with randomly generated forest (ablation) | Ours (Default) · 7-Scenes (average) | frames within 5cm/5deg (%): 93.41% -> 91.06% (93.58% with ICP) | No clear effect |
| replace pre-trained forest with randomly generated forest (ablation) | Ours + ICP · Stanford 4 Scenes (average) | frames within 5cm/5deg (%): 99.85% -> 99.84% | No clear effect |
| pre-train forest on Heads instead of Office (ablation) | Ours (Default) · 7-Scenes (average) | frames within 5cm/5deg (%): 93.41% -> 87.96% | Hurt |
| improvements over original 2017 method (up to 50 modes per leaf) | Ours (2017) · 7-Scenes (average) | frames within 5cm/5deg (%): 84.6% -> 93.41% | Improved |
Closest papers by meaning
- On-The-Fly Adaptation of Regression Forests for Online Camera Relocalisation (CVPR 2017 · similarity 0.95)
- Accelerated Coordinate Encoding: Learning to Relocalize in Minutes Using RGB and Poses (CVPR 2023 · similarity 0.90)
- On the Limits of Pseudo Ground Truth in Visual Camera Re-Localisation (ICCV 2021 · similarity 0.89)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.88)
- Living Scenes: Multi-object Relocalization and Reconstruction in Changing 3D Environments (CVPR 2024 · similarity 0.86)
- 360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries (CVPR 2024 · similarity 0.86)
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.
