GeoPlanAgent — fine-tuned weights

The trained weights behind Plan2Map: A Multimodal Benchmark for Document-Grounded Geospatial Boundary Reconstruction from Planning Records (code · project page).

Both models are trained as 5-fold cross-validation over a shared case→fold split: each fold's copy has a fifth of the benchmark cases held out of its training data, and at inference every case is served by the copy that never saw it during training.

Path	What it is	Performance
sam3_lora/fold_{0..4}/	PEFT LoRA adapters for facebook/sam3, fine-tuned to segment drawn planning boundaries on scanned UK planning maps	0.912 mean pixel IoU
rotation_classifier_kfold/fold_{0..4}/best.pt	ResNet50 (ImageNet-pretrained) fine-tuned to classify scanned-map orientation (0°/90°/180°/270°)	0.981 accuracy (with test-time augmentation)

Usage

These weights are consumed by the GeoPlanAgent pipeline, which handles fold routing, adapter loading, and inference. From the root of a clone of that repository, download them straight into models/:

hf download degenfabian/GeoPlanAgent --include "sam3_lora/*" "rotation_classifier_kfold/*" --local-dir models

The SAM3 base weights are not included — they download from facebook/sam3 (gated; accept Meta's SAM License there) on the pipeline's first run.

Licence

• sam3_lora/ — the adapters are fine-tuned from Meta's SAM 3 and are therefore distributed under the SAM License.
• rotation_classifier_kfold/ — fine-tuned from torchvision's ImageNet-pretrained ResNet50 (BSD-3-Clause); no additional restrictions.

Citation

@misc{Plan2Map2026,
      title={Plan2Map: A Multimodal Benchmark for Document-Grounded Geospatial Boundary Reconstruction from Planning Records}, 
      author={Fabian Degen and Oishi Deb and Jindong Gu and Junchi Yu and Samuele Marro and Philip Torr and Jialin Yu},
      year={2026},
      eprint={2606.02747},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.02747},
}