Research map / 3D vision / Human pose estimation
RelocNet: Continuous Metric Learning Relocalisation using Neural Nets
Vassileios Balntas, Shuda Li, Victor Prisacariu
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 |
|---|---|---|---|
| learn descriptors with frustum overlap loss instead of ImageNet ResNet18 features | retrieval descriptors · 7Scenes heads | relocalisation success rate: 48.6 % -> 70.33% | Improved |
| learn descriptors with frustum overlap loss instead of ImageNet ResNet18 features | retrieval descriptors · 7Scenes stairs | relocalisation success rate: 32.5% -> 62.77% | Improved |
| learn descriptors with frustum overlap loss instead of Places ResNet18 features | retrieval descriptors · 7Scenes fire | relocalisation success rate: 73.1% -> 79.01% | Improved |
| use frustum overlap descriptors instead of PoseNet penultimate features | retrieval descriptors · 7Scenes heads | relocalisation success rate: 29.1% -> 70.33% | Improved |
| use frustum overlap descriptors instead of Ferns | retrieval descriptors · 7Scenes heads | relocalisation success rate: 30.63% -> 70.33% | Improved |
| train retrieval descriptors on more training sequences (10, 50, 100) (ablation) | RelocNet descriptors · 7Scenes heads | relocalisation success rate | Improved |
| fuse k=5 nearest neighbours instead of the single nearest neighbour (ablation) | RelocNet pose estimation · 7Scenes chess | median error: 0.12m, 4.14 -> 0.12m, 3.95 | Improved |
| fuse k=5 nearest neighbours instead of the single nearest neighbour (ablation) | RelocNet pose estimation · 7Scenes heads | median error: 0.14m, 10.5 -> 0.13m, 10.5 | Improved |
| fuse k=5 nearest neighbours instead of the single nearest neighbour (ablation) | RelocNet pose estimation · 7Scenes stairs | median error: 0.28m, 7.53 -> 0.27m, 7.31 | Improved |
| train and test RelocNet on the same 7Scenes sequence instead of ScanNet transfer | RelocNet · 7Scenes chess | median translation error: 0.21m -> 0.12m | Improved |
| use RelocNet trained on ScanNet instead of PoseNet Geometric | RelocNet · 7Scenes stairs | median translation error: 0.35m -> 0.33m; chess 0.13m -> 0.21m | Mixed |
Closest papers by meaning
- Geometric Loss Functions for Camera Pose Regression With Deep Learning (CVPR 2017 · similarity 0.87)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.86)
- Learning Multi-View Camera Relocalization With Graph Neural Networks (CVPR 2020 · similarity 0.86)
- CamNet: Coarse-to-Fine Retrieval for Camera Re-Localization (ICCV 2019 · similarity 0.85)
- Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer (ECCV 2024 · similarity 0.85)
- Dual Pose-invariant Embeddings: Learning Category and Object-specific Discriminative Representations for Recognition and Retrieval (CVPR 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.
