Research map / Continual learning / Domain and prompt-based
ConDo: Continual Domain Expansion for Absolute Pose Regression
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 ConDo, which continually updates a deployed absolute pose regressor using unlabeled inference images by distilling from HLoc (and other scene-agnostic localizers) under a fixed compute budget, evaluated on new benchmarks with scene-condition and pose changes.
Problem
Absolute pose regression (APR) overfits to its training scans and fails catastrophically on novel poses and scene conditions (weather, lighting, season) encountered after deployment, while obtaining new ground-truth labels is costly.
Approach
Continually update the deployed APR with unlabeled inference images, supervising them by distilling predictions from a scene-agnostic localizer (HLoc) while uniformly replaying historical and new data under a limited per-round compute budget; multi-head regression extends it to multiple scenes.
Key results
- On Office Loop inference scans, ConDo reduces Pose-Transformer mean position error from 42.15m to 1.86m and median from 6.12m to 1.50m.
- ConDo with 12 minutes of updates matches Re-train with GT trained for about 5h on Office Loop, a 25x compute/time reduction.
- Replay buffer at 10% of dataset size only slightly degrades ConDo performance on Office Loop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Continually update APR using unlabeled inference data with HLoc distillation | PoseNet and Pose-Transformer · Office Loop and Neighborhood inference scans | median position error: 18.10 -> 2.16; 6.12 -> 1.50; 10.99 -> 1.14; 2.99 -> 0.89 | Helped |
| Continually update APR using unlabeled inference data with HLoc distillation | PoseNet and Pose-Transformer · 7Scenes and Cambridge inference scans | median position error: 0.303 -> 0.080; 0.198 -> 0.065; 1.405 -> 1.052; 1.101 -> 0.696 | Helped |
| Use standard UDA methods RSD or DARE instead of ConDo | Pose-Transformer · Office Loop inference scans | median position error: ConDo 1.50; RSD 6.65; DARE 6.72 | Hurt |
| Add MIC_mae or MIC_moco domain-adaptation supervision (ablation) | Pose-Transformer with ConDo · Office Loop inference scans | median position error: 1.50 -> 4.39; 1.50 -> 4.15 | Hurt |
| Replace HLoc teacher with DM-VIO, ORB-SLAM, or NetVLAD (ablation) | ConDo · Office Loop inference scans | median position error: 1.50 -> 5.21; 1.50 -> 3.39; 1.50 -> 3.00 | Hurt |
| Use a very limited reservoir replay buffer (ablation) | ConDo · Office Loop | median position error: null | Hurt |
| Use multi-head architecture rather than per-scene APR models (ablation) | ConDo · 7Scenes training scans | position/orientation error: null | Hurt |
| Replace the EfficientNet backbone with pre-trained Dino v2 (ablation) | Pose-Transformer · Office Loop inference scans | mean position error: 42.15 -> 21.38 | Helped |
| Combine pre-trained Dino v2 with ConDo (ablation) | Pose-Transformer · Office Loop inference scans | median position error: 1.50 -> 1.02 | Helped |
| Use ConDo rather than equal-budget re-training with GT | Pose-Transformer · Office Loop inference scans, 20h budget | median position error: 1.79 -> 1.50 | No clear effect |
| Use HLoc distillation instead of ground-truth inference labels during re-training (ablation) | PoseNet and Pose-Transformer · 7Scenes training scans | median position error: PN 0.024 vs 0.025; PT 0.026 vs 0.028 | No clear effect |
| Train on additional data with strong pose changes (ablation) | APR · 7Scenes training scans | median position error: null | Hurt |
Surprising findings
- Adding new scans with strong pose changes did not improve accuracy on training scans, even with ground-truth labels (Re-train with GT).
- Standard UDA methods (RSD, DARE, MIC) give minor or negative gains for APR, while distillation from HLoc is far more effective.
Limitations
- Multi-head multi-scene design hurts APR performance, especially for ConDo, on training scans.
- Performance on training scans can decline with pose-change data, indicating APR architectures do not benefit from more diverse data in general.
- Relies on a scene-agnostic teacher (HLoc) with a map built from training data, adding inference-time cost to the server side.
Experimental setup
Datasets: 4Seasons Office Loop 4Seasons Neighborhood 7Scenes Cambridge Landmarks
Benchmarks: Office Loop Neighborhood 7Scenes Cambridge Landmarks
Metrics: median position error (m) median orientation error (deg) mean position error (m) mean orientation error (deg)
Baselines: Train-only Re-train with GT RSD DARE MIC Re-train with HLoc
Builds on: PoseNet Pose-Transformer HLoc MIC reservoir sampling
Closest papers by meaning
- Map-Relative Pose Regression for Visual Re-Localization (CVPR 2024 · similarity 0.90)
- Scene-agnostic Pose Regression for Visual Localization (CVPR 2025 · similarity 0.88)
- Enhancing Visual Localization with Cross-Domain Image Generation (ICML 2025 · similarity 0.87)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.86)
- Neural Refinement for Absolute Pose Regression with Feature Synthesis (CVPR 2024 · similarity 0.86)
- HyperPose: Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset (CVPR 2025 · 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.