HR-APR: APR-agnostic Framework with Uncertainty Estimation and Hierarchical Refinement for Camera Relocalisation

TL;DR

This work addresses the unreliability of absolute pose regressors (APRs) by introducing HR-APR, an APR-agnostic framework that estimates pose reliability through a lightweight pose-based retrieval of training embeddings and uses this uncertainty to guide a NeFeS-based refinement. The uncertainty module is modular and can be plugged into diverse APR architectures, achieving substantial reduction in refinement overhead (up to 27.4% indoors and 15.2% outdoors) while preserving state-of-the-art single-image pose accuracy. Extensive experiments on indoor 7Scenes and outdoor Cambridge Landmarks demonstrate a consistent correlation between uncertainty and pose error, enabling selective refinement and improved robustness without architecture-specific modifications. The approach is efficient, storage-friendly, and suitable for real-time camera relocalisation, highlighting its practical impact for robotics and AR applications.

Abstract

Absolute Pose Regressors (APRs) directly estimate camera poses from monocular images, but their accuracy is unstable for different queries. Uncertainty-aware APRs provide uncertainty information on the estimated pose, alleviating the impact of these unreliable predictions. However, existing uncertainty modelling techniques are often coupled with a specific APR architecture, resulting in suboptimal performance compared to state-of-the-art (SOTA) APR methods. This work introduces a novel APR-agnostic framework, HR-APR, that formulates uncertainty estimation as cosine similarity estimation between the query and database features. It does not rely on or affect APR network architecture, which is flexible and computationally efficient. In addition, we take advantage of the uncertainty for pose refinement to enhance the performance of APR. The extensive experiments demonstrate the effectiveness of our framework, reducing 27.4\% and 15.2\% of computational overhead on the 7Scenes and Cambridge Landmarks datasets while maintaining the SOTA accuracy in single-image APRs.

Figures (7)

• Figure 1: HR-APR: APR-agnostic Framework with Uncertainty Estimation and Hierarchical Refinement.
• Figure 2: The accuracy of APR predictions is highly corresponding to pose variants between query and training images. (a) and (d): Camera ground truth trajectory of Hospital and Church in Cambridge Landmarks dataset kendall2015posenet. Blue: training set; Green: queries in the test set near the training set within 2 meters and 10 degrees; Red: queries in the test set far from the training set. (b), (c), (e), and (f) show that all three different APRs kendall2015posenetshavit2021learningchen2022dfnet have better predictions on test (near) than test (far).
• Figure 3: Uncertainty estimation module.
• Figure 4: Uncertainty evaluation on the 7Scenes and Cambridge Landmarks datasets. Subfigures (a)-(f) show the correlation between similarity threshold $\gamma$ and normalized pose error (0-1). Based on the similarity threshold, uncertain samples with low similarity scores are gradually removed. Subfigures (g)-(i) in the last row show the ratio of retained predictions (%) as threshold increases. We observe that as we remove the samples with low similarity scores the overall error drops indicating a clear correlation between our predictions and the actual inaccurate predictions.
• Figure 5: Plots of translation and rotation errors against the number of iteration for images pass the similarity threshold (hs) and images with low similarity scores (ls) do not pass the similarity threshold on some scenes of 7Scenes. The hs predictions ratio of each scene is provided in Table \ref{['tab:df_level']}. NeFeS$m$ denotes runing the refinement process for $m$ iterations.