CurviWeave: Curvilinear Differential Ribbon-Weaving Network for High-Resolution Remote Sensing Road Extraction
Official model checkpoints for CurviWeave on Massachusetts Roads and DeepGlobe Road Extraction benchmarks.
- Codebase: github.com/Nguyen-Kim-An/CurviWeave
- Datasets: Massachusetts Roads | DeepGlobe Roads
Performance
| Model Checkpoint | Benchmark | Native Tile Size | Global IoU (%) | F1-Score (%) |
|---|---|---|---|---|
curviweave_massachusetts_best.pt |
Massachusetts Roads | 1500 x 1500 | 63.85% | 77.94% |
curviweave_deepglobe_best.pt |
DeepGlobe Roads | 1024 x 1024 | 71.48% | 83.37% |
Quick Start: Download Weights
from huggingface_hub import hf_hub_download
# Download Massachusetts SOTA checkpoint
massa_ckpt = hf_hub_download(
repo_id="k4nnguyen/CurviWeave",
filename="curviweave_massachusetts_best.pt"
)
# Download DeepGlobe SOTA checkpoint
deepglobe_ckpt = hf_hub_download(
repo_id="k4nnguyen/CurviWeave",
filename="curviweave_deepglobe_best.pt"
)
Citation
@article{curviweave2026,
title={CurviWeave: Curvilinear Differential Ribbon-Weaving Network for High-Resolution Remote Sensing Road Extraction},
author={Le-Xuan, Thang and Nguyen-Kim, An and Vu-Manh, Trung and Tran-Ngoc, Hoa},
year={2026}
}