HoliTracer
August 7, 2025 · View on GitHub
Official implementation of the ICCV 2025 paper HoliTracer: Holistic Vectorization of Geographic Objects from Large-Size Remote Sensing Imagery.
Abstract: This paper introduces HoliTracer, the framework designed to holistically extract vectorized geographic objects from large-size RSI. In HoliTracer, we enhance segmentation of large-size RSI using the Context Attention Net (CAN), which employs a local-to-global attention mechanism to capture contextual dependencies. Furthermore, we achieve holistic vectorization through a robust pipeline that leverages the Mask Contour Reformer (MCR) to reconstruct polygons and the Polygon Sequence Tracer (PST) to trace vertices. Extensive experiments on large-size RSI datasets, including buildings, water bodies, and roads, demonstrate that HoliTracer outperforms state-of-the-art methods.
1. Installation
Requirements
- OS: Linux distribution, our paper experiments are based on Ubuntu 22.04
- Hardware: A GPU with at least 24GB of memory and CUDA support is needed for training (8GB is enough for inference). Our experiments were performed on NVIDIA A100 GPUs with 40GB of memory.
Setup Instructions
git clone https://github.com/vvangfaye/HoliTracer.git
cd HoliTracer
pip/conda install torch torchvision # our paper experiments are based on pytorch 2.5.1
pip install -r requirements.txt # install other dependencies
# install pycocotools with holitracer compatible version.
git clone https://github.com/vvangfaye/cocoapi-holi.git
cd cocoapi-holi/PythonAPI && python setup.py install
# install holitracer
cd ../../ && pip install -e . # install holitracer with editable mode
2. Datasets
| Dataset Name | Image Size | Spatial Resolution | Images | Train/Val/Test | Download Link |
|---|---|---|---|---|---|
| WHU-building | 10,000 × 10,000 | 0.075 m | 400 | 320 / 40 / 40 | Google Drive |
| GLH-water | 12,800 × 12,800 | 0.3 m | 250 | 200 / 25 / 25 | Google Drive |
| VHR-road | 12,500 × 12,500 | 0.2 m | 208 | 166 / 21 / 21 | Google Drive |
Download the datasets from the provided links and extract them to the data/datasets directory.
3. Model Zoo
Pre-trained models and performance metrics:
| Dataset | PoLiS ↓ | CIoU | AP | APs | APm | APl | IoU | F1 | Download Link |
|---|---|---|---|---|---|---|---|---|---|
| WHU-building | 3.63 | 82.30 | 61.07 | 40.37 | 80.30 | 60.00 | 91.60 | 95.41 | Google Drive |
| GLH-water | 81.87 | 59.24 | 20.84 | 19.88 | 38.77 | 72.29 | 85.68 | 91.51 | Google Drive |
| VHR-road | 134.13 | 6.10 | 1.58 | 0.08 | 0.40 | 3.99 | 46.48 | 60.63 | Google Drive |
Download the pre-trained models from the provided links and extract them to the data/models directory.
4. Inference and Visualization
Run the demo file to get started:
- demo.py and geo_demo.py(TIFF and Shapefile support): Includes examples for inference and visualization.
Or run the python notebook
5. Training
Training scripts and instructions are available in:
Refer to this file for detailed steps to train HoliTracer on target dataset.
Contact
If you have any questions about it, please let me know. (Create an 🐛 issue or 📧 email: wangfaye@whu.edu.cn)
We are developing a unified vectorization framework for remote sensing imagery in EarthVec, and we are happy to collaborate with you.