Source-linked AI summary
AirAlign: Geometry-Aware Relative Pose Alignment for UAV Last-Meter Navigation
Jinyi Zhou, Shuo Feng, Yufei Wu, Piji Li
TL;DR
UAV last-meter navigation needs precise relative pose alignment despite severe viewpoint and appearance changes. AirAlign uses geometry-aware RGB image-pair features with scene-disjoint fold-based model selection and ensembling, achieving a high PairUAV challenge ranking.
Problem
UAV last-meter navigation requires accurate final-approach pose alignment, but viewpoint and appearance variations make relative pose recovery from RGB image pairs challenging.
Method
AirAlign uses a pretrained visual geometry reconstruction model for geometry-aware source-target features and averages predictions from models selected through scene-disjoint cross-validation.
Results
AirAlign achieves a high ranking on the PairUAV challenge test set, with a final score of 0.002790 and sixth place on the final leaderboard.
Takeaways & Limitations
The reported challenge performance and ablation studies demonstrate AirAlign's effectiveness for RGB-only UAV image-pair relative pose alignment.
Abstract
from arXiv · showhide
Unmanned aerial vehicle (UAV) navigation in modern low-altitude environments requires more accurate pose alignment in the final approach stage for target information acquisition or manipulation, making "last-meter" navigation increasingly important. However, severe viewpoint and appearance variations make this task challenging. To tackle this problem, we propose AirAlign, a framework for RGB-only image-pair relative pose alignment for UAVs. AirAlign uses a pretrained visual geometry reconstruction model as the backbone to extract geometry-aware features from source-target image pairs. In addition, to better utilize the limited training data, we split the training set into multiple scene-disjoint folds for unseen cross-validation and model selection. During inference, the predictions of the selected models are averaged to form the ensemble output of the overall framework. Experiments on the PairUAV challenge at the ACMMM 2026 Workshop on UAVs in Multimedia demonstrate the effectiveness and robustness of our method, while comprehensive ablation studies validate the contribution of each component.
1 Introduction
UAV last-meter navigation requires precise final-approach pose alignment, but viewpoint changes and limited sensing make image-pair geometry difficult to recover. AirAlign addresses this with geometry-aware RGB features, scene-disjoint cross-validation, and ensemble prediction.
- Last-meter navigation requires accurate pose alignment near targets because position or heading errors can affect landing safety, delivery accuracy, and close-range observation.
- Different altitudes and headings create substantial appearance and spatial-arrangement changes between source and target images, challenging pose recovery from appearance or global semantics.
- Existing UAV visual-navigation studies mainly address place-level localization or retrieval, whereas target localization alone is insufficient for final-stage geometric pose alignment.
- AirAlign formulates RGB-only UAV navigation as predicting relative translation and heading from a source-target image pair.
- The framework uses scene-disjoint folds for cross-validation, independently trains multiple models, and averages selected-model predictions during inference.
- AirAlign combines a pretrained visual geometry reconstruction model with fold-based ensemble prediction, and its PairUAV challenge performance and ablations support the framework's effectiveness.
2 Related Work
Prior visual-geometry and aerial-navigation research provides tools for geometric reconstruction, correspondence, exploration, and localization. However, the cited navigation methods mainly target reaching the vicinity, leaving fine-grained relative pose alignment as AirAlign's focus.
- Traditional SfM and MVS recover 3D structure through feature matching and stereo reconstruction, while feedforward geometry models predict geometric representations more directly.
- MASt3R extends visual-geometry modeling with image-pair correspondence grounded in 3D geometry rather than appearance similarity alone.
- Goal-oriented aerial-navigation systems such as UAV-ON, CityNav, and OctMem-Agent use language instructions, memory, or exploration for target-directed navigation.
- For image-goal navigation, ANWM uses a world model for planning, while GeoExplorer uses curiosity-driven active exploration.
- These navigation methods mainly reach the target vicinity, whereas AirAlign predicts relative translation and heading from paired UAV views for fine-grained pose alignment.
3 Methodology
AirAlign predicts relative translation and heading from RGB source-target image pairs using geometry-aware features, task-specific heads, scene-disjoint cross-validation, and ensemble inference. Its architecture combines camera-pose and 3D-point-map representations, while training and inference strategies improve model selection and prediction robustness.
- 3.2 Geometry-Aware Feature Extraction: AirAlign extracts geometry-aware representations from source-target RGB images using a pretrained visual geometry reconstruction backbone.The backbone includes geometry feature extraction and dedicated camera-pose and 3D-point-map decoders.
- 3.2 Geometry-Aware Feature Extraction: The backbone outputs camera-pose and point-map features, which are concatenated to form pair-level geometry-aware representations.Concatenation occurs along the feature dimension.
- 3.3 Relative Translation and Heading Prediction: Two three-layer MLP heads predict relative translation and heading; translation uses camera-pose and point-map features, while heading uses camera-pose features.Heading is represented as a two-dimensional unit direction vector, and translation outputs are mapped to a normalized range with a sigmoid.
- 3.4 Training Strategy and Model Ensemble: Training uses scene-disjoint folds, rotating each fold as an unseen validation set to select the checkpoint with the lowest final score.The remaining folds train the model for each round, and the pretrained backbone is frozen while prediction heads are optimized.
- 3.4 Training Strategy and Model Ensemble: The framework trains heading and translation with smooth L1-based objectives, including normalized direction, normalized translation, and relative error terms.The overall loss weights relative angle and relative distance error losses through λ_ang and λ_dist.
- 3.4 Training Strategy and Model Ensemble: Inference averages predictions from multiple trained models, averaging heading direction vectors before recovering the final angle and averaging translation predictions.The ensemble contains M models.
4 Experiments
AirAlign is evaluated on PairUAV using relative-error metrics and ablations of its fold strategy, auxiliary losses, and input features. The framework achieves a final score of 0.002790 and ranks 6th on the challenge leaderboard.
- Overall Performance: AirAlign achieves a final score of 0.002790, with distance relative error 0.003175 and angle relative error 0.002405.The result is reported on the PairUAV challenge test set.
- Ablation Study: The 5-fold setting achieves the best final score of 0.002790, outperforming 3-fold and 7-fold settings at 0.004239 and 0.003883, respectively.The ablation indicates that more folds do not necessarily improve performance.
- Ablation Study: Removing the corresponding auxiliary relative losses increases distance error by 0.008840 and angle error by 0.001196, respectively.These results support the contribution of both auxiliary relative losses to final performance.
- Ablation Study: Adding point map features to the heading head increases angle error to 0.003691, while removing camera pose features raises distance error to 0.011359.Heading prediction mainly relies on camera pose features, whereas translation prediction benefits from both camera pose and point map features.
5 Conclusion
AirAlign is a relative pose alignment framework for UAV last-meter navigation based on a pretrained visual geometry reconstruction model. It uses scene-disjoint folds for training and selection, averages selected-model predictions, and achieves a high PairUAV challenge ranking.
- Conclusion: AirAlign uses a pretrained visual geometry reconstruction model for relative pose alignment in UAV last-meter navigation.The framework targets image-pair relative pose alignment during the final approach phase.
- Conclusion: Scene-disjoint training folds support model training and selection, while selected-model predictions are averaged during inference.The conclusion identifies this fold-based ensemble procedure as part of the framework.
- Conclusion: AirAlign achieves a high ranking in the PairUAV challenge at the ACMMM 2026 Workshop on UAVs in Multimedia.The conclusion presents this ranking as evidence of the framework’s effectiveness.