Technical Summary: AirAlign

Problem Statement

The paper addresses the challenge of "last-meter" navigation for Unmanned Aerial Vehicles (UAVs) in low-altitude environments. While UAVs are increasingly used for tasks like infrastructure inspection and logistics, the final approach phase requires precise pose alignment for target interaction, which is difficult to achieve with standard GNSS. The core difficulty lies in the significant viewpoint and appearance variations between a source image (captured from a distance) and a target image (captured close to the goal). These variations, combined with the constraint of using only RGB images without auxiliary depth sensors, make recovering the spatial relationship and relative pose (translation and heading) between the two views highly challenging. Existing solutions often focus on coarse place-level retrieval rather than the fine-grained geometric alignment necessary for safe landing or manipulation.

Methodology

The authors propose AirAlign, a framework for RGB-only relative pose alignment that leverages geometry-aware features rather than relying solely on appearance textures. The methodology consists of three primary components:

1. Geometry-Aware Feature Extraction

Instead of training a model from scratch, AirAlign utilizes a pretrained visual geometry reconstruction model, specifically $\pi^3$ [17], as its backbone. This model, which incorporates a DINOv2 encoder and alternating view-wise/global self-attention layers, is designed to predict geometric representations from image pairs.

• Feature Decoding: The backbone processes a source-target image pair to generate hidden geometry features. These are then passed through two dedicated decoders: a camera pose decoder and a 3D point map decoder.
• Representation: The system concatenates the features from both the source and target images to form pair-level representations ($f_{cam}$ and $f_{pts}$) that encode the relative spatial relationships.

2. Task-Specific Prediction Heads

The original output heads of the $\pi^3$ model are replaced with two lightweight Multi-Layer Perceptrons (MLPs) to predict the relative pose:

• Heading Prediction: This head takes only the camera pose features ($f_{cam}$) as input. It outputs a 2D unit direction vector to represent the relative angle, avoiding the complexity of direct angle regression.
• Translation Prediction: This head utilizes both camera pose features ($f_{cam}$) and point map features ($f_{pts}$) to capture comprehensive geometric information. It outputs a normalized translation vector using a sigmoid activation function.

3. Training Strategy and Ensemble Inference

To maximize the utility of the limited training data and improve robustness, the authors employ a scene-disjoint cross-validation strategy:

• Folding: The training set is split into multiple non-overlapping scene folds (e.g., 5 folds).
• Training & Selection: Multiple models are trained independently, where each fold serves as a validation set once while the others train the model. The best checkpoint from each round is selected based on the lowest validation score.
• Ensemble: During inference, the predictions from all selected models are averaged. The heading vectors are averaged and converted back to an angle, while the normalized translations are averaged and denormalized to form the final ensemble output.

Key Contributions

The paper outlines three main contributions:

1. AirAlign Framework: A novel image-pair relative pose alignment framework specifically designed for UAV "last-meter" navigation, utilizing a pretrained visual geometry reconstruction model to enable fine-grained pose estimation.
2. Scene-Disjoint Ensemble Strategy: A training methodology that splits the dataset into scene-disjoint folds for cross-validation and model selection, followed by an ensemble inference process to enhance robustness.
3. Empirical Validation: Demonstration of the framework's effectiveness through strong performance on the PairUAV challenge test set and comprehensive ablation studies validating the necessity of each component (folding strategy, auxiliary losses, and feature inputs).

Experimental Results

The method was evaluated on the PairUAV challenge at the ACMMM 2026 Workshop on UAVs in Multimedia. The evaluation metrics included distance relative error, angle relative error, and a combined score.

• Performance: AirAlign achieved a final score of 0.002790, ranking 6th on the leaderboard.
• Comparison: This result significantly outperforms the official baseline score of 0.633957.
• Ablation Studies:
  • Fold Count: A 5-fold ensemble yielded the best results (0.002790), outperforming 3-fold (0.004239) and 7-fold (0.003883) settings, suggesting an optimal balance between model diversity and validation scene coverage.
  • Auxiliary Losses: Removing the relative angle loss increased the angle error by ~0.0012, while removing the relative distance loss increased the distance error by ~0.0088, confirming the value of these auxiliary terms.
  • Feature Inputs: The study revealed that heading prediction relies primarily on camera pose features (adding point map features degraded performance), whereas translation prediction benefits significantly from both camera pose and point map features.

Significance and Claims

The paper claims that AirAlign effectively addresses the limitations of existing retrieval-based approaches by explicitly modeling 3D geometric relationships through a pretrained backbone. By focusing on the "last-meter" phase, the work provides a solution for scenarios where GNSS is insufficient and fine-grained alignment is critical for safety and task success. The authors assert that their scene-disjoint ensemble strategy and geometry-aware feature extraction are key factors in achieving robust performance under severe viewpoint variations, as evidenced by their high ranking in the challenge. The work is presented as a practical step toward enabling autonomous UAVs to perform precise interactions in complex, low-altitude environments using only standard RGB imagery.