Similarity-based selection
Geometric constraints are complemented by visual similarity to retain a compact set of compatible image pairs.
Abstract
Bundle adjustment of multi-date satellite imagery is particularly challenging because seasonal, illumination, and land-cover changes can make conventional feature correspondence pipelines unreliable. This work extends RPC bundle adjustment with season-invariant correspondences and a pair-selection strategy designed to avoid visually incompatible image pairs.
The proposed pipeline incorporates learned detector-based matching methods, including combinations such as SuperPoint and LightGlue, and augments geometric pair-selection constraints with a visual-similarity criterion. These changes improve the robustness and efficiency of RPC refinement on heterogeneous multi-date image collections while retaining corrected camera models in standard RPC format.
Method at a glance
Learned correspondences
The matching stage supports learned detector-based pipelines such as SuperPoint combined with LightGlue.
Robust RPC refinement
Season-invariant feature tracks support bundle adjustment across images acquired at different dates.
Software release
sat-bundleadjust v2.0.0
This major release extends the correspondence and pair-selection components of the original pipeline for multi-date imagery.
- Learned detector-based matching pipelines, including SuperPoint + LightGlue.
- Similarity-based pair selection that retains a subset of visually compatible image pairs.
- Improved robustness to seasonal and acquisition-dependent appearance changes.
- Continued support for RPC bundle adjustment and RPC-format camera export.
Citation
@inproceedings{mari2026robust,
title = {Robust RPC Bundle Adjustment for Multi-Date Satellite
Imagery with Season-Invariant Correspondences},
author = {Marí, Roger and Masquil, Elías and
Bou, Xavier and Ehret, Thibaud and
Facciolo, Gabriele},
booktitle = {ECCV Workshops},
year = {2026}
}