LoRetta
Official inference checkpoint for LoRetta: A Foundation Model and Extensive Dataset for Global-Scale Remote Sensing Dense Image Matching.
- Project page: https://www.siweiyu.com/work/loretta/
- Paper: https://arxiv.org/abs/2608.04106
- Code: https://github.com/yusiwei77-star/LoRetta
- Authors: Siwei Yu, Han Guo, Zhenwei Shi, and Zhengxia Zou
- Institution: LEVIR Lab, Beihang University
- File:
loretta.pth
- Format: canonical tensor-only PyTorch state dictionary
- State-dictionary entries: 807
- File size: 1,294,957,584 bytes
- SHA-256:
09a502056b671d4e07819f454f96eb385ebe605a186ee1b059ce2447d8fb602e
Model description
LoRetta is an inference model for dense image matching in remote-sensing imagery. It follows a localization-and-registration design: a matchability-aware module estimates the overlapping region and an affine prior, then a guided coarse-to-fine registration module predicts residual dense correspondence in the aligned frame. The released package uses a pinned, locally vendored DINOv3 ViT-L/16 backbone architecture and does not fetch executable model code at runtime.
Installation and usage examples are available in the
GitHub repository.
Intended use
LoRetta is intended for research on remote-sensing image registration, cross-view matching, geometric alignment, and geolocalization. Users must have the right to process their imagery and should independently validate the results.
Do not treat model output as verified evidence or a final decision in safety-critical, legal, surveillance, property, emergency-response, or other high-impact settings.
Input and output assumptions
- The reference and sensed images are independently resized to 512 × 512; their original aspect ratios are not preserved.
- The sensed image is expected to have a ground sampling distance no greater than the reference image and a ground footprint no larger than the reference image.
- Outputs include a dense warp field and a per-pixel matchability estimate. Optional geometry estimation derives a homography from predicted matches using RANSAC.
- Very large aspect-ratio changes, weak overlap, severe appearance change, clouds, repetitive texture, or violations of the scale/footprint assumptions can degrade results.
Data overview
LoRetta was developed with LEVIR-GM, a multi-temporal optical remote-sensing matching benchmark described by the authors as containing 103K aligned pairs and 827K augmented pairs across six continents, five years, and spatial resolutions from 0.5 to 1024 metres. The
paper is public on arXiv; the dataset, detailed sampling documentation, training code, and evaluation code are not yet public. Users should not infer demographic, geographic, sensor, season, or land-cover coverage beyond the published description.
Evaluation
On LEVIR-GM, the authors report:
- 83.3% AUC, 1.6 points higher than RoMa v2.
- PCK improvements of 6.5 points at 1 pixel and 8.2 points at 2 pixels.
- 47.8% lower inference latency than RoMa v2.
PCK measures how many matches fall within a pixel error threshold, while AUC summarizes accuracy across thresholds. Latency depends on the implementation and hardware. These figures are reported in the paper; the evaluation code is not yet public.
Limitations and risks
- Repetitive or texture-poor areas, clouds, weak overlap, and strong appearance changes can produce plausible but incorrect matches.
- Errors can propagate into incorrect transforms or locations. Preserve uncertainty, reject low-confidence results, and independently verify important outputs.
- Performance may degrade outside the stated input assumptions or on underrepresented sensors, regions, seasons, and land-cover types. Resizing to 512 × 512 can distort extreme aspect ratios. Users are responsible for lawful and ethical use.
License and prohibited uses
This checkpoint incorporates DINOv3 parameters and is distributed under the
DINOv3 License. Use and redistribution must comply with that agreement. In particular, the license requires compliance with applicable laws, privacy/data-protection rules, and trade controls, and prohibits covered end uses including military or warfare purposes, nuclear applications, espionage, and the development or use of guns or illegal weapons. This summary is informational only; the license text controls.
LoRetta-authored source code is distributed separately under Apache License 2.0. Third-party source components retain their respective licenses. See the code repository's THIRD_PARTY_NOTICES.md for provenance and complete notices.
If results produced with this checkpoint are submitted for publication, the DINOv3 License requires acknowledgment of the DINO Materials. Please also cite LoRetta as shown below.
Citation
1@article{yu2026loretta,
2 title = {LoRetta: A Foundation Model and Extensive Dataset for
3 Global-Scale Remote Sensing Dense Image Matching},
4 author = {Yu, Siwei and Guo, Han and Shi, Zhenwei and Zou, Zhengxia},
5 journal = {arXiv preprint},
6 year = {2026}
7}
Security contact
Please report suspected vulnerabilities privately to contact@siweiyu.com. Do not include access tokens, private imagery, credentials, or other sensitive data in the initial report.
Copyright 2026 LEVIR Lab, Beihang University.
