Dense 3D Displacement Estimation for Landslide Monitoring via Fusion of TLS Point Clouds and Embedded RGB Images

Zhaoyi Wang · Jemil Avers Butt · Shengyu Huang · Tomislav Medic · Andreas Wieser

Geosensors and Engineering Geodesy (GSEG) Lab, Institute of Geodesy and Photogrammetry (IGP), ETH Zürich

---

This repository contains official implementations of our series of work in TLS-based landslide monitoring. ⭐ Star us if you find it useful!

• 📷 RGB-Guided Dense 3D Displacement Estimation in TLS-Based Geomonitoring (ISPRS Geospatial Week, 2025)

• 🏔️ Dense 3D Displacement Estimation via Fusion of TLS Point Clouds and Embedded RGB Images (JAG, 2026)

The main branch contains the official implementations of both works and baselines (Piecewise ICP, F2S3). For the code specific to the RGB-Guided approach, switch to the rgb-guided-only branch.

---

📚 Table of Contents

1. Installation
2. Example data and weights
3. Run
4. TODO list
5. Citation

🛠️ Installation

We test the code with Python 3.8 and PyTorch 2.4.1 + CUDA 12.4 (cu124). If you use a different CUDA / PyTorch version, please update the installation commands accordingly.

git clone --recursive https://github.com/gseg-ethz/fusion4landslide.git
cd fusion4landslide

# create conda env + install torch/pyg/cpp dependencies
source install.sh

📁 Example data and weights

We provide our collected Rockfall Simulator dataset: 📦 Download from Hugging Face. This dataset is partially used in our RGB-Guided paper. Please refer to the paper for a detailed data description and feel free to explore the full dataset for your own research.

The pretrained weights used can be downloaded from here, please download weight files and save them under './weights/'.

🚀 Run

Before running, modify the corresponding config file in /configs/test/[method]_[dataset].yaml, and update any necessary parameters in main_[method].py. Then, run the main pipeline using:

python main_fusion.py          # run for Point cloud+RGB fusion (fusion) or Point cloud-only (fusion_3d) methods
python main_rgb_guided.py      # run for RGB-Guided method
python main_f2s3.py            # run for F2S3 baseline
python main_piecewise_icp.py   # run for Piecewise ICP baseline

📌 TODO list

• Add sample data and RGB-Guided scripts.
• ⏳ Add baselines for quantitative and qualitative comparison.
• ⏳ Release full RGB-3D fusion pipeline upon paper acceptance.
• ⏳ Extend to photogrammetric 3D point clouds (SfM/MVS).
• ⏳ Provide Pythonic version of point cloud tiling.

🤝 Acknowledgements

We gratefully acknowledge the following open-source projects that contributed to this work:

• Superpoint segmentation (original · customized): Used for generating hierarchical patches in our pipeline.
• Supervoxel segmentatition (original · customized): Incorporated for single patch generation.
• F2S3: 3D geometry-based displacement estimation baseline.
• Efficient LoFTR (original · customized): Used for semi-dense image pixel matching in our RGB-Guided method, and as the image-matching module in our fusion approach.
• RoMA (original · customized): Integrated for dense image pixel matching.

🔗 Relevant projects

• py4dgeo: A repository that contains implementations of M3C2 and its variants.
• iecepy4D: A Python package for image-based glacier monitoring.
• Piecewise-ICP: A 4D point cloud registration method.

🤗 Citation

If our work helps your research, please consider citing:

@article{wang2026fusion4landslide,
  title={Dense 3D displacement estimation for landslide monitoring via fusion of TLS point clouds and embedded RGB images},
  author={Wang, Zhaoyi and Butt, Jemil Avers and Huang, Shengyu and Medić, Tomislav and Wieser, Andreas},
  journal={International Journal of Applied Earth Observation and Geoinformation},
  volume={146},
  pages={105093},
  year={2026},
  issn={1569-8432},
}

@article{wang2025RGB4landslide,
  title={An approach for rgb-guided dense 3d displacement estimation in tls-based geomonitoring},
  author={Wang, Zhaoyi and Butt, Jemil Avers and Huang, Shengyu and Meyer, Nicholas and Medić, Tomislav and Wieser, Andreas},
  journal={ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
  pages={953--960},
  year={2025},
}