Are Pretrained Image Matchers Good Enough for SAR–Optical Satellite Registration?

Isaac Corley · Alex Stoken · Gabriele Berton

CVPR 2026 Image Matching Workshop

We evaluate 24 pretrained image matchers zero-shot on SAR–optical satellite registration across SpaceNet9, SRIF, and SARptical — no fine-tuning, no domain adaptation. The best matchers (RoMa, XoFTR) reach 3.0 px mean tie-point error on SpaceNet9, and we show that protocol choices alone (tile size, geometry model, normalization) can shift accuracy by up to 33× for a single matcher. Full results and analysis are on the project page.

Install

git clone https://github.com/isaaccorley/rsim.git
cd rsim
make install

Requires Python 3.12+ and a CUDA GPU (evaluated on a single RTX 3090). make install runs uv sync --all-groups and installs pre-commit hooks.

Download data

SpaceNet9: bash scripts/download-spacenet9.sh

SRIF: Download from LJY-RS/SRIF and unpack to data/srif/.

SARptical: Download from the TUM SiPEO group and unpack to data/sarptical/patch_SAR_OPT_SQUARE/.

Reproduce the paper

uv run python scripts/run_protocol_sweep_top_matchers.py   --device cuda
uv run python scripts/run_extended_transfer_ablations.py   --device cuda
uv run python scripts/run_sarptical_spacenet_methods.py
uv run python scripts/generate_paper_assets.py

CSVs land in outputs/; tables and figures land in paper/tables/ and paper/figures/.

Citation

@inproceedings{corley2026rsim,
  title     = {Are Pretrained Image Matchers Good Enough for {SAR}--Optical Satellite Registration?},
  author    = {Corley, Isaac and Stoken, Alex and Berton, Gabriele},
  booktitle = {CVPR Workshop on Image Matching},
  year      = {2026}
}

License

Released under the Apache 2.0 License by Isaac Corley.